Warehouse, operatives and driver optimization process in a parcel delivery company

Master’s thesis Warehouse, operatives and driver optimization process in a parcel delivery company Autor: Antonio Castaño Blanco Director: Ginés Alarcón Submit date: September 2020 Escola Tècnica Superior d’Enginyeria Industrial de Barcelona Page 2 Thesis Index INDEX _______________________________________________________ 2 INDEX OF FIGURES____________________________________________ 5 1. INTRODUCTION ___________________________________________ 7 1.1. Objectives ...................................................................................................... 8 1.2. Scope of the project ....................................................................................... 9 2. WAREHOUSE DESIGN ____________________________________ 10 2.1. Picker ........................................................................................................... 10 2.2. States of a parcel ......................................................................................... 10 2.2.1. Draft ................................................................................................................. 11 2.2.2. Pending ............................................................................................................ 11 2.2.3. On course ........................................................................................................ 11 2.2.4. Picked-up ......................................................................................................... 11 2.2.5. Delivered .......................................................................................................... 11 2.2.6. Return in progress ........................................................................................... 11 2.2.7. Undelivered ...................................................................................................... 12 2.3. Routing ......................................................................................................... 12 2.4. Sorting .......................................................................................................... 12 2.5. Roundcheck ................................................................................................. 12 2.6. Dispatch ....................................................................................................... 13 2.7. Order’s process ............................................................................................ 13 3. OPERATIONS ____________________________________________ 15 3.1. Dispatching .................................................................................................. 15 3.2. Inventory ...................................................................................................... 15 4. CURRENT WAREHOUSE LAYOUT __________________________ 16 4.1. Organizational policies ................................................................................. 18 4.2. WH Design problem ..................................................................................... 19 4.2.1. Strategic level .................................................................................................. 19 4.2.1.1. Process flow design ............................................................................. 19 4.2.2. Tactical level .................................................................................................... 20 4.2.3. Operational level .............................................................................................. 20 5. OPERATIONS’ PARAMETERS ______________________________ 21 5.1. Volume related ............................................................................................. 21 Warehouse, operatives and driver optimization process in a parcel delivery company Page. 3 5.2. Time related.................................................................................................. 22 5.3. Labour related .............................................................................................. 22 6. COST ANALYSIS _________________________________________ 25 6.1. Fixed costs ................................................................................................... 25 6.2. Variable costs ............................................................................................... 25 7. WH CURRENT CONSTRAINTS ______________________________ 26 7.1. External factors ............................................................................................. 26 8. WH CURRENT MARGINS __________________________________ 27 8.1. Fleet parameters .......................................................................................... 27 8.2. Gross margins .............................................................................................. 27 8.2.1. GM1 ................................................................................................................ 31 8.2.2. GM2 ................................................................................................................ 31 8.2.3. GM .................................................................................................................. 32 9. WAREHOUSE VOLUME ___________________________________ 34 9.1.1. LSTM Network ................................................................................................ 37 9.1.2. Architecture ..................................................................................................... 37 10. LATE TRUCK ARRIVALS __________________________________ 41 10.1. Number of parcels processed depending on the late truck arrival time ....... 41 11. WAREHOUSE AUTOMATION _______________________________ 45 11.1. IOT Systems automation .............................................................................. 46 11.1.1. Check point ..................................................................................................... 47 11.1.2. Equipment ....................................................................................................... 49 11.1.3. Cost analysis ................................................................................................... 51 11.2. Conveyor ...................................................................................................... 53 11.2.1. Cost analysis ................................................................................................... 57 11.3. Next steps..................................................................................................... 58 12. CONCLUSIONS __________________________________________ 59 13. COSTS _________________________________________________ 60 14. ANNEX _________________________________________________ 62 14.1. Neuronal network code ................................................................................ 62 14.1.1. Libraries .......................................................................................................... 62 14.1.2. Creating methods ............................................................................................ 62 14.1.2.1. make_into_t_series .......................................................................... 62 Page 4 Thesis 14.1.2.2. create_neuronal_model.................................................................... 63 14.1.3. Input data processing ....................................................................................... 63 14.1.4. Neuronal network implementation .................................................................... 64 15. BIBLIOGRAPHY __________________________________________ 66 15.1. Complementary bibliography ....................................................................... 66 Warehouse, operatives and driver optimization process in a parcel delivery company Page. 5 Index of figures Figure 1: Parcel delivery market size worldwide (Source: statista) ......................................... 8 Figure 2: Order's process diagram ....................................................................................... 14 Figure 3: WH Area cadastre ................................................................................................. 17 Figure 4: WH current layout .................................................................................................. 17 Figure 5: WH parameters. Volume related ........................................................................... 21 Figure 6: WH Parameters. Time related ............................................................................... 22 Figure 7: Labour costs .......................................................................................................... 23 Figure 8: WH Parameters. Driver related ............................................................................. 24 Figure 9: Monthly margins .................................................................................................... 29 Figure 10: GM1 vs Sales ...................................................................................................... 31 Figure 11: GM2 vs Sales ...................................................................................................... 32 Figure 12: Total GM vs Sales ............................................................................................... 32 Figure 13: Barcelona daily volume March – April ................................................................. 34 Figure 14: Volume distribution .............................................................................................. 35 Figure 15: Neuronal network architecture ............................................................................. 36 Figure 16: Training process validation .................................................................................. 38 Figure 17: Training function at epoch 1 ................................................................................ 39 Figure 18: Training function at epoch 1000 .......................................................................... 39 Figure 19: Forecast values compared to the real ones ......................................................... 39 Figure 20: Table with transformed values ............................................................................. 40 Figure 21: Dispatching timings ............................................................................................. 42 Page. 6 Thesis Figure 22: Quantity of parcels that can be processed depending on the truck arrival ........... 43 Figure 23: New sorting process ............................................................................................ 48 Figure 24: Scanner layout .................................................................................................... 49 Figure 25: Scanner model .................................................................................................... 50 Figure 26: ESP32 ................................................................................................................. 51 Figure 27: Sorting productivity .............................................................................................. 52 Figure 28: New sorting vs Old sorting................................................................................... 53 Figure 29: Sorter flow chart .................................................................................................. 54 Figure 30: Cross belt sorter .................................................................................................. 55 Figure 31: Sorting process ................................................................................................... 56 Figure 32: Volume division by the size of the parcels ........................................................... 57 Figure 33: Large parcels operators assigned ....................................................................... 57 Figure 34: Labour costs ........................................................................................................ 60 Figure 35: Material costs ...................................................................................................... 61 Figure 36:Total cost .............................................................................................................. 61 Warehouse, operatives and driver optimization process in a parcel delivery company Page. 7 1. Introduction Ecommerce is a growing industry. Since Amazon delivered its first book in 1995, to encompass the 95% of all purchases by 2040 according to the marketing agency 99 firms. But that is still far from the current situation. With a growth of 19% over 2019, there is a lot still to be discovered, which leaves a door open for entrepreneurs to find a place in the market. Such growth is leading to a turn on events that will change (and it is currently doing so) the global economic markets. Distribution centres are a fundamental part for this market’s growth, being (at least nowadays) a necessary piece on the whole process. From the moment the client orders something until it gets delivered there are two main processes regarding logistics: • The retailer’s facilities: Where the product must be located and loaded into a truck to be sent to the distribution centre. • Distribution centre: Most commonly known as warehouse. Its purpose is to receive and process all the freight received by many retailers, and prepare it for the last mile delivery, where vans follow a route delivering all the parcels received from the retailers. A distribution centre must be able to face the yearly increase of demand by also increasing its capacity and efficiency, maintaining or decreasing its costs. The competitiveness of a sector such as distribution centres, where its yearly revenues have increased a 47.6% in the last ten years (figure 1), comes along with a profit subjected to both the final client’s high demands regarding fast and personalized delivery and the retailer’s volume of parcels fluctuations. This narrows the profit margins left for the warehouse company. How to enhance it while fulfilling the mentioned requirements is key to evolve and adapt within the market’s growth. The growth factor might also cause problems if some of the main variables of parcel delivery such as parcel control, warehouse capacity and operatives’ availability in a short period of time are not faced in time. Page. 8 Thesis Figure 1: Parcel delivery market size worldwide (Source: statista) In addition, European Union green industry policies such as industrial renewal, competitiveness, sustainability and eco-efficiency must be considered as milestones for the future of delivery industry. 1.1. Objectives The main objective of this project is to analyse the delivery company performance regarding the dispatchment process so to study how the current processes and profit margins interrelate, to finally improve those margins to a minimum of the average of the market, by optimizing the processes considered as productivity/cost bottleneck. This analysis will be carried out from the moment the warehouse receives a parcel to the instant the parcel is delivered. Once the analysis is performed and conclusions are taken from it, a series of proposals will be analysed so to be able to enhance the warehouse productivity ratios at a feasible cost. To do so, two different approaches will be taken: • Provide a warehouse automation and control system. • Improving the already existing processes. But first, in order to be able to reach those approaches, we must perform a complete study of Warehouse, operatives and driver optimization process in a parcel delivery company Page. 9 the company, its warehouse layout and all its processes. The aim of this project is to find a new warehouse model that optimizes the parcel processing, reducing the costs and improving the service offered to the final client. After the project is implemented, the warehouse will be an adaptive environment that can successfully re define its layout within short notice. Through developing a new model, we will be able to keep up with the company's growth rate, and to change variable costs to fixed costs independently to the volume of parcels. 1.2. Scope of the project It is crucial to correctly establish the scope so its boundaries are set and we can focus on what is within the scope. In Scope: The parameters considered as inside the scope are mainly everything that takes place or operates in the warehouse: • Labour related: From the cost of the operators to setting the number of workers. • Processes: They will be analysed and changed in case the overall productivity can be enhanced. • Layout: The way the warehouse is distributed can be essential to be able to process and control as much volume as possible within the lowest cost. Out of scope: These parameters are considered as fixed, and its change won’t even be considered: • Out of warehouse: not a single process nor parameter taking place out of the warehouse will be considered in this project. • Warehouse facilities: changing the warehouse to a new one that fits better with the processes is out of the scope, since due to a contract doing so is not possible for several years. • Retailer: Anything related to changing the conditions with the retailers such finding an agreement to improve the conditions or getting new clients. • Labour cost: The costs associated to labour work are fixed. Page. 16 Thesis 4. Current warehouse layout Once the processes have been laid out, the current layout can be explained and discussed. The second most relevant process regarding costs is the one involving parcel sorting and dispatching, both activities taking place in the warehouse. To set the proper KPIs and to measure variables such as operative’s productivity, average and peak volume and retailer truck’s arrival time is essential so the warehouse performance can be studied and improved. A proper performance comes along with the correct number of workers and timetable so to enhance the margins obtained from delivering a parcel. But before getting a close look at the warehouse’s processes, it is important to study the current warehouse size, distribution and location. The current Barcelona warehouse is located in Parc Logistic, a logistic area in Zona Franca with a total of 120,000 m2 of facilities aimed for the logistic and e-commerce sector. It is a one floor building of 1920 m2, distributed in 6 different areas: • Load/Onload dock: It is a six-dock area, where the trucks load and unload their freight. Pre-sorting process also takes place in this area. • Sorting area: Where the parcels are stored before dispatching. They are sorted by the route the driver will follow to deliver them. • Parking area: Area where drivers park their vehicles to intake the parcels. • Office: Mainly for routing and solving issues that take place during dispatching. • Parcels to return area. • Restroom. Warehouse, operatives and driver optimization process in a parcel delivery company Page. 17 Figure 3: WH Area cadastre Figure 4: WH current layout Page. 18 Thesis WH Operation areas: 1. Load/Unload and dispatchment area. 2. Pre-sorting boxes. 3. Sorting pallets. 4. To be returned orders. 4.1. Organizational policies WH organizational policies’ aim are to define the warehouse layout according to the following standards: • Organizational policies: Formed by the assignment policies establishing the allocation of trucks to docks and other vehicles such as vans. • Storage policies: Irrelevant for our case of study since the freight manage in a parcel delivery warehouse isn’t generally stored for more than one day. • Order picking policies: They refer to the picking and storing hierarchy. • Operators assignment policies: Related to the number of operatives working in each shift. The facility has the workers, delivery drivers and truck drivers entrance located in the same side. This affects the WH layout process flow, forcing it to be bidirectional and limiting the organizational policies regarding the vehicle assignment policies: The freight arrives at the same place where the drivers come and pick up the routes. This forces operation related to trucks and related to vans to be coordinated so no in load/outload freight and dispatchment take place at the same time. The storage area is at the bottom of the warehouse, and the parking area is just between the storage area and the in load/unload area. Once the freight is out loaded from the truck, picking and sorting take place sequentially, first picking the parcels and then sorting them into their routes. Because the big volume fluctuations the warehouse process on a daily basis, the operatives, which come from an external agency, are called upon the forecasted weekly volume. Warehouse, operatives and driver optimization process in a parcel delivery company Page. 19 4.2. WH Design problem There are three main approaches regarding the warehouse design problem. These approaches depend on each other according to how it affects the WH performance in terms of time, being the longer-term decision a constrain for the lower ones. These approaches are classified into strategic, tactical and operational. 4.2.1. Strategic level Where long term impact decisions are considered. Topics such as process flow design and selecting the machinery WH systems are being considered taking into account WH technical capabilities and economic considerations. 4.2.1.1. Process flow design Regarding process flow design, both the input and the output material and products must be described. Figure 5 shows and order’s flow, since it leaves the retailer’s facilities until it is delivered. The input are parcels coming from different retailers, whose delivery time interval is set by the costumer. In the other hand, the output are routes of parcels that are designed basing on the delivery postal code and the delivery time interval so when a driver is dispatching the route, the dispatching order and delivery location make sense regarding their time table and location. Figure 5: Orders flow chart Page. 20 Thesis So, the main design problems here these three: • Are the truck arrival times standardized? • How are the parcels sorted after the trucks unload? • How are the routes distributed? This question is out of the scope of the thesis. In terms of choosing the WH systems, a business case regarding the possibility of automating the warehouse entirely or just some selected process must be carried out. Currently, there are not many automated processes regarding parcel processing and delivery, apart from the order getting a picker assigned. This will be further discussed in the following sections. The required equipment of a distribution centre doesn’t go beyond the standard nonautomated warehouse tools. 4.2.2. Tactical level After setting the process flow and the warehousing systems, we have to dimension the WH different zones (sorting, storing, dispatching, etc) and its resources such as the work items, the operatives and the processing systems (in case there are any). In order to set this up correctly, the average and peak volume processed on a daily basis must be established, and size parameters such as the width of the alleys. This is crucial to minimize the over costs that might be applied due to an over dimension of the WH size or operatives’ quantity. A cost analysis will be carried out later in order to see which are the costs and if they fit the gross margins set by the company. 4.2.3. Operational level Carried out within the constraints set by strategic and tactical levels. It studies the assignment and control problems of operatives and equipment, in addition to the routing system. Warehouse, operatives and driver optimization process in a parcel delivery company Page. 21 5. Operations’ parameters Before starting the analysis that will lead to future conclusions and decisions, we have to define and quantify the variables that are relevant during the operations. These variables may constitute a wide variety and quantity. It has major importance which variables are selected to measure the warehouse activity throughput. Most of the parameters change within time, so we have set the month of march to set our variables, although in a real approach they should be analysed after the last updated data inputs. 5.1. Volume related • Volume: 3500 parcels/day. Although volume is a fluctuating variable that depends on the time of the year, the average volume avoiding high volume periods (for this case, a different study must be carried out) has been calculated taking into account the 2020 orders received by retailers except Black Friday and Christmas days. 90% of these parcels are to be delivered during the morning shift, whereas the other 10% belongs to the afternoon shift. • Parcels per route: This is a constant value set at 40, constraint by the van capacity and the uncertainty of the parcels’ volume, which forces the van to leave an extra space in case a route’s orders volume exceed the average. • Number of drivers: 88 This parameter comes from the two mentioned above, being the number of drivers required for delivery, the volume divided by the parcels per route (since the route refers to the number of parcels a driver delivers in a shift sequentially). 79 of those drivers work in in the morning shift (as mentioned, 90% of the parcels are delivered in the morning) and 9 in the afternoon. Volume related Quantity Units Volume 3500 parcels Driver Capacity 40 parcels/route Morning volume 90% Afternoon volume 10% Figure 5: WH parameters. Volume related Page. 22 Thesis 5.2. Time related • Sorting productivity: 120 parcels/hour/operative. This parameter is out of the thesis’ scope. It measures the number of parcels per hour that an operative should process under a correct performance during sorting. • Roundcheck productivity: 540 parcels/hour/operative. It is a much faster process compared to sorting productivity, since this phase of the process consists only in double checking that all parcels are put in their route. Although it is way quicker than sorting, it is not as relevant, it is performed due to the high volume of wrongly sorted parcels. Later on, alternatives to this process will be further discussed. To both productivities, there is an extra margin that should be applied when calculating the timings, due to operatives’ extra time that might be spent moving around the WH or solving daily incidences. TIME RELATED Quantity Unit Sorting productivity 120 parcels/h/op Sorting capacity 30 s/parcel Roundcheck productivity 540 parcels/h/op Margin for incidences 20% Figure 6: WH Parameters. Time related • Already set timings o Truck arrivals: 4:30 am. This is the starting point for the operatives. Once the trucks arrive, the operative onload the freight and start sorting. The time this process takes depends on the number of operatives working at the same time. o Dispatching start time: 8:50 am. At this time, drivers arrive at the WH to load their routes and start delivering. 5.3. Labour related To be able to fit the timings, which are already restricted by the truck arrivals and the earliest time range delivery, at 9 am, we have to fix the proper number of operatives for each operation before dispatching: Sorting and roundcheck. The current number of operatives per operation depends only on the volume of parcels, regardless of the difference of productivity that both processes (sorting and roundcheck present) The truck arrivals and dispatching start time sets a restriction for sorting and roundcheck timings, hence, also for the number of operatives. The margin between both processes is of Warehouse, operatives and driver optimization process in a parcel delivery company Page. 23 4 hours and 20 minutes. So, the total hours along with the number of operators must fit that space. Sorting Roundcheck n Op Time total [h] Cost [€] 5 5,3 1,4 6,5 400 6 4,4 1,2 5,5 406 7 3,8 1,0 5 431 8 3,3 0,9 4 394 9 2,9 0,8 3,5 387 10 2,6 0,7 3,5 431 11 2,4 0,6 3 406 12 2,2 0,6 3 443 13 2,0 0,5 2,5 400 14 1,9 0,5 2,5 431 15 1,8 0,5 2 369 Time [h] Figure 7: Labour costs Figure 7 shows the total hours of labor required depending on the number of operators working at the operative, along with the total time and the daily cost attached. The time total column is calculated by rounding up the sum of both the sorting and roundcheck time since the shifts’ sensitivity can’t be lower than half an hour. Another constraint set by the temporary agencies is that the operators working hours cannot be under three hours and a half. Currently, the number of operators used for an average morning shift is 10, which is not the optimal quantity since the number of operators can be lowered to 9 while meeting the requirement mentioned above. In case the volume forecast indicates the contrary, this quantity must change accordingly (by just changing the volume field in the excel file). However, taking into account the productivity of sorting and roundcheck, it might be worth taking into consideration adjusting the number of operators for each operation. The costs attached to operators may vary depending on the shift, being the normal shift salary around 12€/h to 15€/h during night shift and extra hours. Parameters related to the drivers are described in Figure 8. Page. 24 Thesis Number of drivers 88 drivers Morning 79 drivers Afternoon 9 drivers Cycle time 15 min/van Capacity 4,0 van/h Max vans 20 vans Total Capacity 80 vans/h Load 59 min Driver related Figure 8: WH Parameters. Driver related The number of drivers is directly related to the volume and the driver capacity (Figure 5) Out of the total volume, the 10% is usually delivered in the afternoon whereas the other 90% is delivered in the morning, the number of drivers operating in both shifts are similarly distributed. The total capacity indicates the number of vans that can be dispatched per hour taking into account the maximum number of vans that can fit into the parking area. The load time has been calculated taking into account the morning shift numbers of drivers for it is the bottle neck of the process. Warehouse, operatives and driver optimization process in a parcel delivery company Page. 25 6. Cost analysis To be able to improve the shipping process, it is necessary to have a clear view of the costs, so to focus on the ones that seem higher than expected. Costs are divided into fixed and variable. Fixed costs are those that remain constant independently of the volume, drivers, etc. On the other hand, variable costs change as one of the previously described parameters change. 6.1. Fixed costs These types of costs are divided by the days the WH is operating, so the margins can be studied on a daily basis. This study was carried out during the month of May, with a total of 24 operating days. • Warehouse rent: 1480 €/month, a daily cost of 61.66 €/day. 6.2. Variable costs • Drivers: The highest of the variable costs. It depends directly upon the volume of orders to be delivered that day. As stated in figure 5, every driver may take up to 40-45 parcels per route, so the drivers needed for a day are exactly the volume of orders divided by the driver’s parcel capacity. Drivers currently come from an external agency, and the quantity required is forecasted on a weekly basis. Because they have attached the highest cost, they have the highest responsibility into delivering the parcels in the time range they have assigned. Every fleet of drivers has its own conditions attached, being the most relevant the price paid to the fleet per parcel delivered within its time slot. • Operators: Also, from external agencies. Similar to drivers, the cost is directly related to the volume and to the operative’s productivity, which has been set to 100 parcels/hour/operative for sorting plus presorting and of 540 parcels/hour/operative. It is also essential to know the trucks’ arrival time so to establish the available time for the operation to be done. • Parcel pick-ups: All the costs related to picking up the parcels, whether from the retailer’s facilities or directly from the store, are charged in here. This cost varies within the number of trailers required, which change depending the volume of Page. 32 Thesis 66% 68% 70% 72% 74% 76% 78% 80% 82% 84% 86% 0 5.000 10.000 15.000 20.000 25.000 % GM2 € GM2 Sales GM2 GM2% Figure 11: GM2 vs Sales GM2 presents a higher margin’s ratio than GM1, almost doubling it with an 84.5% at its peak and an average of 80%. This shows that the greatest losses come from the last mile delivery phase. However, although GM1 is way lower, GM2 present more factors and a higher ratio of improvement in terms of productivity and process design. 8.2.3. GM Figure 12 shows the total gross margin, subtracting GM2 to GM1, and comparing it to the daily sales. The overall gross margin is of 47400 €, which compared to the total sales makes for a profit of a 13%. 0% 5% 10% 15% 20% 25% 30% 35% 0 5.000 10.000 15.000 20.000 25.000 % GM € GM Sales GM GM% Figure 12: Total GM vs Sales Warehouse, operatives and driver optimization process in a parcel delivery company Page. 33 So, how are the numbers compared other models in the market? According to Capgemini institute study, carried out in 2018, the overall average cost, which accounts for warehouse processing and last-mile delivery accounts for a 28.7%, two times the company’s overall gross margin. It is fundamental to re define the shipping model and design a strategy in order to gradually cut distance between both numbers and, once achieve the 28.7% of profit find new ways to increase it. In the following sections, different new projects are approached to reach the established margins, starting for the processes related to GM2 and then following to the ones involving the Last-Mile delivery, always having in mind the main goal, which is to get to the average gross margin per parcel within a time period of two years. Page. 34 Thesis 9. Warehouse volume The warehouse labor system is not fixed, it varies directly within the daily demand. Basing upon the expected volume, the warehouse asks the work agencies in case of the operatives and to the fleet companies for the drivers for a specific quantity of workers. A shift on volume has to rapidly translate into a shift on labor demand. An accurate volume prediction is crucial to identify and apply those shifts fast enough without over nor under scaling, which means an increase on the daily gross margin per parcel. To have a clear view on demand, historical volume data will come in handy. Unfortunately, for companies with a short run other options should be considered and studied. For this case of study, the volume dataset contains 52 days, as indicated in Figure 13. Figure 13: Barcelona daily volume March – April To get a proper forecast it is crucial that it follows a trend or a pattern, this way the forecasting process can identify and predict basing upon them. In this case, there is a pattern that repeats no matter the higher or lower volume trend, that is, two peaks and one valley every week. The peaks belong to the beginning and the end of the week, being one fixed on Friday and another distributed between Monday and Tuesday depending on the week. The valley is always on Saturday, where volume week after week reaches the lowest in the week. This may change within time and then so will do the algorithm estimating the volume to process. Figure 14 shows the distribution of daily volume depending on the day of the week, being 1 Monday and 6 Saturday (the warehouse remains close on Sundays). The graphic shows Warehouse, operatives and driver optimization process in a parcel delivery company Page. 35 that, in spite of the dispersion that specific days show, around three per day, there is also a pattern trend following that dispersion. Wednesday is the most irregular day of the week, which can be observed both through figure 5 and 6. 0 200 400 600 800 1000 1200 1400 1600 1800 2000 01234567 Volume Day of the week Volume weekly distribution 1 2 3 4 Figure 14: Volume distribution There are many ways to approach a solution, from solutions that only take into account the weekly volume and it trend towards past weeks to considering the holidays and other special days such as sales in order to apply a scale factor. It remains clear the need of a forecast whose results vary upon time, and that takes into account factors such as the week day and the trend of the curve. Taking into account these requirements, the solution we’ve come up with is building a forecast model using neuronal networks that are time related. Artificial neuronal networks are mathematical models that emulate how the brain behaves and process data. They are especially useful for non-linear dataset interpretation; however, they may also be used for linear data sets [1]. A neuronal network is composed different types of neurons: • Predictors: They are the input data that is to be processed by the network. They form the bottom layer of the network. Every predictor has a weight associated to them, depending on how their presence reflects on the output. The weights are selected in Page. 36 Thesis the neural network framework using a “learning algorithm” that minimises a “cost function” such as the MSE [1]. • Hidden neurons: Forming the intermediate layer. They are used for nonlinear data and the number of hidden neurons may vary upon the nonlinearity degree. Time related inputs don’t require hidden neurons since they are (in most of the cases) linear. • Output layer: It processes all the predictors and outputs the forecast value. Figure 15: Neuronal network architecture For time series, the previous dates may be used as input data in form of lagged vales in an autoregressive model just to get the following day as an output of all the lagged inputs taking into account the season ability of the data along with the trend line. There are some factors that affect how the forecast’s accuracy [2]: • Trend: Defined by the long-term gradual increase of volume. • Cycle: Defined by how the volume varies following a pattern. Warehouse, operatives and driver optimization process in a parcel delivery company Page. 37 • Seasonality: It controls the short-term fluctuations of volume. • Error: Similar to noise on a signal, a time series may have fluctuations that cannot be explained by the model. 9.1.1. LSTM Network To be able to control these factors and build a model that outputs a proper volume estimation, we have decided to use Long Short-Term Memory networks (LSTM). LSTM process the input data as sequential, so all the inputs are treated as dependent one from each other, another feature of these network is the capacity of remembering long-term sequences being able to associate and process them in order to output a trustworthy output. The neuronal network has been built up using Python programming environment, since it can provide many useful tools for designing and implementing the LSTM model. In order to prove that the model estimations are valid a train and test method has been developed. Data will be split into two groups, so the neuronal network is trained using the first group and then it can be tested by forecasting the values from the second group. This way, we can compare the forecast to the real values registered and decide whether the network has been properly trained or not. 9.1.2. Architecture Every neuronal network must have set its proper architecture, so the forecast is as accurate as possible. These processes can follow some parameters to get the optimal one, although many times the best option is acquired by trying different parameters. The following parameters have been set by a combination of both. • Inputs: 6. According to the number of days the warehouse is operating. • Output: 1, for the number of days we want to forecast. • Hidden neurons: 10. To estimate the number of hidden neurons the following formula has been applied. • Activation function: tanh. Used to introduce non linearities to the output network [3]. • Epochs: 1000. It defines the number times that the learning algorithm will work Page. 38 Thesis through the entire training dataset. For the program to run we have made used of the libraries Keras and Pandas, which are especially designed for neuronal network processing. Figure 16 shows the results from the training process. The red points are the current values while the green ones are the trained ones, which are trying to approach the real values. The closer they get, the better the train. This is not a closed training, but a continuous one. Within time, the dataset will increase and the neuronal network’s train will improve. Figure 16: Training process validation From Figures 17 and 18 the difference between the beginning and the end of the training process is noticeable, from the first epoch error and mean square values to the last one. The last epoch presents pretty low loss and mse values. Warehouse, operatives and driver optimization process in a parcel delivery company Page. 39 Figure 17: Training function at epoch 1 Figure 18: Training function at epoch 1000 Now, results from figures 17 and 18 will be compared: • Loss: 92.03% reduction. • MSE: 99.02% reduction. The epochs effect is clear, both the loss and the minimum square error are hugely reduced. 0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000 05/03/2020 10/03/2020 15/03/2020 20/03/2020 25/03/2020 30/03/2020 04/04/2020 09/04/2020 14/04/2020 19/04/2020 Real vs Forecasted volume Volume Volume Forecasted Figure 19: Forecast values compared to the real ones Real and trained trend lines are compared in Figure 19, where the orange line approach the blue one, although the difference between values could be stretched (the bigger the dataset the better) the trend is correct in every value. This proves the accuracy of the forecasting. Once the network is trained, now the volume is to be forecasted. For the network to work with the tanh activator, the dataset is changed to oscillate between -1 and 1, and the dataset transformed into this structure: Page. 40 Thesis Figure 20: Table with transformed values This way, the forecasted value will be the outcome of var1(t+1) at the last row. The forecasted value for the current dataset is 4123, corresponding to the 20th of April 2020. Warehouse, operatives and driver optimization process in a parcel delivery company Page. 41 10. Late truck arrivals A truck getting later than expected, and therefore unloading outside the allocated time slot affects directly to the routing and sorting process, leading to a late dispatchment, and consequently to delayed or failed delivery. So, this raises the question of how to proceed when a truck arrives later than expected, which outcome is better in terms of profit and in terms of client satisfaction. There are three possible approaches to this question: • To wait for the delayed truck: This solution enhances the probability of delivering more parcels for that day. The whole process would be delayed, waiting for the delayed trucks to be sorted and moved to their route areas. Every route containing a parcel from the delayed truck would end up delaying the delivery of all the parcels, which would reduce the margins per parcel, for a parcel delivered out of its time range has a penalization from the retailer. • To deliver without the delayed parcels. In this case, the non-delivered parcels due to the truck’s late arrival are for the retailer to blame, so no penalization is imputed to the company. However, this approach harms the client directly, and the company’s reputation. • To set a quantity of parcels that may be sorted and dispatched from the late trucks, depending the total volume and the arrival time. This proposal is halfway between the other two, delivering as many parcels as possible without compromising the margins, the time slots nor the company’s reputation. 10.1. Number of parcels processed depending on the late truck arrival time To set the number of parcels that may be processed we have to set a number of operators, since both data are correlated. The values this study has been carried out upon are the ones depicted previously in the parameters’ section. The number of parcels that can be processed without interfering with the last mile delivery are directly related to the timings of the whole process. The first time slots attempts are set at 10:20, so this is the first constraint regarding the delivery. It has also been stated the average time taken at arriving to the first destination, 20 minutes and an extra time for any kind of incidence (Figure 20). Page. 48 Thesis Figure 23: New sorting process Warehouse, operatives and driver optimization process in a parcel delivery company Page. 49 When the double scan is carried out, the parcel is added to the route’s check point. Once all the parcels are sorted, the operatives can check out at the check point route list that all its parcels were sorted into the box. Figure 23 shows how the scanner would be placed relative to the pallet. Figure 24: Scanner layout 11.1.2. Equipment To implement this service, different possibilities have been studied. The main points taken into account when looking for solutions were the following: • Robustness: the solution implemented must be robust, so after installing the whole layout there are no problems regarding the service. • Affordable: automating a process is usually y high cost investment. In order to get profit within a low ROI, the solution must come with an affordable price or else it won’t be implemented. The chosen scanner model is YK-EP3000 2d barcode scanner module Wiegand, y fixed horizontal scanner in the shape of a box so it can be placed by the route box. It is a simple device that apart from the barcode reader has a USB port so it can be connected to another device to exchange data Although the average number of routes is 88, the purchased quantity of scanners will be oversized by a 15% in case the volume increases or a scanner gets broken, meaning Page. 50 Thesis that the number of scanners to buy is 101. The cost for each scanner has been set to 33.15€. Figure 25: Scanner model A device able to collect the information from the scanner to process it and communicate with the collector devices is required. A good solution for this, that also meets the requirements set for the scanner is the ESP32 controller. The ESP32 is currently one of the most complete low-cost micro controller solutions in the market. Here are some of its features: • Small size (25.5 x 18.0 x 2.8mm). • Low cost: each microcontroller has a cost of 3.06 €. • WIFI and Bluetooth low energy integration: Allowing the device to communicate with the server and directly with the operator phone. • Low energy consuming: Due to the Ultra-Low Processor, the device can enter the Sleep Mode, which allows the device to be off (light sleep mode), waiting for the scanner to detect a barcode. Warehouse, operatives and driver optimization process in a parcel delivery company Page. 51 Figure 26: ESP32 Integrating both devices so they are connected to each other is not enough. To be able to share and store the information obtained by the barcode scanner it is necessary set the machine to machine (M2M) connection so to establish the IoT environment. To connect an object to the IoT, it is necessary to have internet connection. For this project, WIFI has been chosen as the network connection method due to the routes changing nature (the routes layout may vary depending upon the volume to process). The object must connect to a server in order to set the communication between devices. All the changes affecting the objects connected to the IoT must be updated to the server. This allows real time communication between all devices. Data transmission between server and client is carried out through HTTP Requests (Hypertext Transfer Protocol Request), which follows the request-answer protocol between devices. These communications must be encrypted by an SSL protocol (Super Socket Layer), so no external device can have access to the data. 11.1.3. Cost analysis Although the double scan decreases the sorting productivity per parcel, it erases the round check operative. The next steps focus on studying the throughput of this new process compared to the original one. Page. 52 Thesis Because the original sorting process already included time for moving the scanned order to its route box, the time the has to be added to the sorting productivity is only the one involving the second scan. Scanning at the fixed route-box scanner takes an extra of three seconds per parcel, which are accountable for scanning it and waiting for the sound alarm approving or denying the check point pairing. The new sorting numbers would be the following: TIME RELATED Quantity Unit Sorting capacity 27 s/parcel Sorting productivity 133 parcels/h/op Figure 27: Sorting productivity This changes the timings related to the total hours required depending on the operators along with the processes timing restrictions set to be able to deliver the parcels on time. Figure _ shows the comparison between the new and the old process in terms of number of hours required depending on the operators working. To fill the requirements set from the temporary agencies establishing that the workers shift must be above 3.5 hours. Old sorting required 3.5 hours and 9 workers to fulfil the hours needed for processing the orders, whereas the new sorting needs just 8 workers and 3.5 hours to accomplish the job at a much lower cost. Overall, the savings related to removing the roundcheck process are of 44 € per day, which represents a 11% less than the old sorting. Warehouse, operatives and driver optimization process in a parcel delivery company Page. 53 n Op Time [h] Cost [€] Time [h] Cost [€] 5 6,00 369 400 6,50 6 5,00 369 406 5,50 7 4,00 344 431 5,00 8 3,50 344 394 4,00 9 3,00 332 387 3,50 10 3,00 369 431 3,50 11 2,50 338 406 3,00 12 2,50 369 443 3,00 13 2,00 320 400 2,50 14 2,00 344 431 2,50 15 2,00 369 369 2,00 Old Sorting New Sorting Figure 28: New sorting vs Old sorting 11.2. Conveyor The next logical point automation wise after enabling the IoT communication in the warehouse is to study the possibility of installing a conveyor capable to do the sorting into route automatically. By installing a conveyor most of the dispatching process would be automated, with the exception of the trucks’ unload, printing the company’s barcode and moving the parcels to the sorter. The conveyor would carry out the following tasks: • Scan the parcel. • Sort the parcel into the assigned route box. The parcels’ flow through the conveyor can be observed in figure 29. The orders are scanned at the beginning, so they are output when they pass by the outlet leading to the route-box the parcel has been attached to. Page. 54 Thesis Figure 29: Sorter flow chart There are currently many conveyor systems in the market, so we must choose the one that fits better our requirements without getting an over-featured one that may increase the price and reduce our margins. The solution we have come up with uses the Cross-belt Sortation Conveyor for automating the warehouse’s layout. Cross Belt conveyors used to sort parcels, apparel, and small items, at high speed [4]. The parcels are carried out in individual carts, which are small bi-directional conveyor belts. Each spot is a belt that gets activated when it lines up with the parcel’s route-box destination. This method main advantage is the individual place for each parcel, so the orders will be separated and can’t get mixed nor missorted. When the tray reaches the destination position, the mechanical belt is activated and the parcel ejected to its box. Thanks to the control system, the sorter is able to output the parcel in the assigned route-box. Warehouse, operatives and driver optimization process in a parcel delivery company Page. 55 Figure 30: Cross belt sorter The new sortation process is described in Figure 31, the tasks colored in yellow are the one manually performed whereas the green ones are processed by the sorter. Parcels are sorted automatically. This new layout reduces the tasks carried out by labor, so a new cost analysis must be performed to establish the viability of the automation project. In addition, it increases the total productivity, due to the sorters’ throughput, which is of 15 pieces/s. This conveyor may be used to process small-medium parcels and bags, the 80% of the daily freight received by the retailers (this is assured by a deal with the retailers), the other 20% must be handed by the operatives. Page. 56 Thesis Figure 31: Sorting process Warehouse, operatives and driver optimization process in a parcel delivery company Page. 57 11.2.1. Cost analysis To make a decision on whether to automate the warehouse by setting a conveyor we must compare the both the current layout and the check-point based model to the conveyor to decide which of the three is the best in terms of fixed and variable costs, labor costs and productivity. The cross-belt cost is 9500 €/set, which consists on the whole layout formed by: 2 belts with 45 routes each, which makes 90 routes in total. Total Small/medium Large Volume 3500 2520 630 Parcel size Figure 32: Volume division by the size of the parcels Two operatives will carry out the sorting process for the large parcels and another two will be in charge of watching out for the cross-belt sorter not to fail and placing the parcels into the sorter. Large volume Sorting productivity n OP 630 parcels 150 parcels/op/h 2 Figure 33: Large parcels operators assigned The gross yearly savings exceed the machine acquisition, so the sorter conveyor is worth a purchase for the warehouse automation, which will lead to a higher productivity and lower costs. Checkpoint 8 Conveyor 2 Diff 6 x n hours 3.5 x op cost/hour 12.3 Daily savings 258.3 €/day x operating days 24 x 12 months 12 Yearly savings 74390 €/year Page. 64 Thesis # ensure all data is float values = values.astype('float32') # normalize features scaler = MinMaxScaler(feature_range=(-1, 1)) values = values.reshape(-1, 1) scaled = scaler.fit_transform(values) # frame as supervised learning reframed = make_into_t_series(scaled, INPUTS, 1) 14.1.4. Neuronal network implementation In this part of the code the training and testing is defined and implemented. The train days are designed as the 80% of the total of the days in the dataset. #Neuronal network definition values = reframed.values n_train_days = int(len(values)*0.8 train = values[:n_train_days, :] test = values[n_train_days:, :] # split into input and outputs x_train, y_train = train[:, :-1], train[:, -1] x_val, y_val = test[:, :-1], test[:, -1] # reshape input to be 3D [samples, timesteps, features] x_train = x_train.reshape((x_train.shape[0], 1, x_train.shape[1])) x_val = x_val.reshape((x_val.shape[0], 1, x_val.shape[1])) print(x_train.shape, y_train.shape, x_val.shape, y_val.shape) The number of epochs is set to 1000, and the model is created and the training started. #set number of epochs EPOCHS=1000 model = create_neuronal_model() history=model.fit(x_train,y_train,epochs=EPOCHS,validation_data=(x_val,y_ val),batch_size=None) Here, the results from the training are tested and compare with the real ones to check if the training was successful. #Test the training results=model.predict(x_val) plt.scatter(range(len(y_val)),y_val,c='g') plt.scatter(range(len(results)),results,c='r') plt.title('validate') plt.show() values = df[-18:].values values = values.astype('float32') Warehouse, operatives and driver optimization process in a parcel delivery company Page. 65 # normalize features values=values.reshape(-1, 1) scaled = scaler.fit_transform(values) reframed = make_into_t_series(scaled, INPUTS, 1) reframed.drop(reframed.columns[[0]], axis=1, inplace=True) values = reframed.values x_test = values[len(values)-1:, :] x_test = x_test.reshape((x_test.shape[0], 1, x_test.shape[1])) y = model.predict(x_test) inverted = scaler.inverse_transform(y) reframed = make_into_t_series(scaled, INPUTS, 1) reframed.drop(reframed.columns[[5]], axis=1, inplace=True) values = reframed.values x_test = values[len(values)-1:, :] x_test = x_test.reshape((x_test.shape[0], 1, x_test.shape[1])) forecast_value = model.predict(x_test) forecasted_val = scaler.inverse_transform(forecast_value) #PRINTS THE RESULT print(forecasted_val) Pàg. 66 Memòria 15. Bibliography [1] Hyndman, R.J & Athanasopoulos, G. (2020, 08 16). Forecasting: principles and practice, 2nd edition. Retrieved from https://otexts.com/fpp2/nnetar.html [2] Lazzeri, F. (2020, 07 03). Medium. Retrieved from https://medium.com/ [3] Jain, S. (2020, 07 22). Activation functions in Neural Networks. Retrieved from https://www.geeksforgeeks.org/activation-functions-neural-networks [4] Cross Belt Sorters. (2020, 09 01). Retrieved from https://www.bastiansolutions.com/ [5] Brownlee, J. (2020, July 24). Machine Learning Maestry. Retrieved from https://machinelearningmastery.com/ 15.1. Complementary bibliography Brownlee, J. (2020, July 30). Multi-Step LSTM Time Series Forecasting Models for Power Usage. Retrieved from https://machinelearningmastery.com/how-to-develop-lstm- models-for-multi-step-time-series-forecasting-of-household-power-consumption/ Keras. (2020, July 19). Retrieved from https://keras.io/api/layers/recurrent_layers/lstm/ N. Kolban, Kolban’s Book on ESP32, USA: Leanpub, 2017.