Design of a 150kW transformer with integrated series inductance for future aerospace applications by using artificial NN

Design of a 150kW transformer with integrated series inductance for future aerospace applications by using artificial NN Alberto Delgado1 , Andres Ferrer1 and Miroslav Vasic1 1 Universidad Politécnica de MadridCentro de Electrónica Industrial, España Corresponding author: Alberto Delgado, [email protected] Speaker: Miroslav Vasic, [email protected] Abstract The future of transportation hinges on the electrification of all vehicles. This article presents a novel design methodology for high power transformers in DC-DC converters dedicated to propulsion, specifically targeting voltage levels between 600 and 1.2 kV for next-generation regional electric aircraft. Utilizing an artificial neural network (ANN), we establish an initial design benchmark for general specifications. Subsequently, we refined this design based on detailed specifications and established design rules, achieving a transformer with integrated series inductance of 20 kW/L integrating the liquid cooling. This work paves the way for more efficient and powerful electric propulsion systems in aviation. 1 Introduction The trend in the last years in commercial aircraft mobility has been the evolution from traditional hydraulic, pneumatic systems to a more electric aircraft in which these systems are mostly replaced by electrical ones reducing the weight, volume and increasing the reliability. Even some technologies such as the Electric Green Taxiing System (EGTS) have been presented to further reduce fuel consumption. The EGTS is a system that allows the transport of the aircraft with electrical motors to drive the wheels avoiding the use of fuel or external vehicles to move the aircraft while it is on the ground. Some of the B737 aircraft have been updated with this system [1][2]. Furthermore, platforms such as B787 and A350 have proved the feasibility of weight reduction by replacing most of the old mechanical, pneumatic and hydraulic systems with electrical ones [3]. In this context, the Advisory Council of Aeronautics Research in Europe (ACARE), set several goals to be achieved by 2050. One of these goals is to achieve zero CO2 emissions in intra-Flight EU and flights departing from Europe. The demonstration of successful MEA systems, and ACARE goals by 2050 has paved the way to go one step further and present projects such as HECATE by Clean Aviation which aims to develop a hybrid electrical regional aircraft [4]. As can be thought, power electronics take a fundamental role in this project. There are three main supply rails/buses: Very High Voltage Direct Current (KHVDC), High Voltage Direct Current (HVDC) and Low Voltage Direct Current (LVDC). DC-DC converters are necessary to convert and distribute power, resulting in challenges related to power density, weight, and thermal management. Furthermore, MEA and HECATE project face knew challenges associated with electromagnetic interferences and their effect on aircraft communications that must be analyzed to develop robust systems [5]. To be able to reduce the size and increase the power density in voltage level adaptation and/or provide isolation, Solid State Transformer (SST) is used in AC grids that are in the low frequency range. An SST also known as power electronics transformer consists essentially of a converter that is directly connected to the low frequency AC grid that is used to pass from the low frequency range (Hz range) to medium frequency (kHz range), then goes through the MFT transformer that provides isolation and apply a voltage ratio depending on turns ratio and finally, another converter to go back to the low frequency range. The power converters and the MFT operate at frequency in the range of few to several kHz [6][7]. This approach provides a major advantage over simply using a low frequency AC transformer that is the increase of the power density of the transformer that leads to an obvious reduction in size and weight and new possibilities such as power flow control, voltage and current regulation and the implementation in not only AC grids but also DC/DC buses. In the literature, it is possible to find different methodologies to design high-power transformers. In [8], a 150 kW transformer is designed and optimized to minimize both temperature rise and volume, resulting in a 13.3 kW/L transformer. In [9], a 100 kW transformer with high insulation is designed, resulting in 20.6 kW/L. In [10], a 100 kW transformer has been obtained through Pareto front optimization with an initialization vector with the required electrical parameters, resulting in a 17.7 kW/L transformer. In [11], a 40kW transformer is designed with a similar methodology to the one presented in [10], yielding an 8.7 kW/L transformer. Finally, in [12], a 200 kW transformer is designed with the focus on thermal management and resulted in a 18.5 kW/L transformer. However, in this paper, using the proposed design techniques, a transformer with integrated series inductance of approximately 23 kW/L has been achieved, including the water-cooling heatsink. Section 2 introduces the aNN used for optimization, explaining its inputs, outputs, and geometry generation. Section 3 outlines the design rules applied for parameter optimization, geometry definition, and conductor optimization based on the neural network’s results. Finally, Section 4 presents experimental results to validate the prototype. 2 Artificial Neural Network Construction Artificial neural networks can approximate complex functions, which makes them suitable for solving optimization problems. It is a deterministic system, with predefined and fixed inputs and outputs. Therefore, the use of these techniques for transformer optimization has to be defined from the beginning. It is necessary to select the type of core, where the windings are positioned and the parameters that define both geometries. In this case, the inputs of the artificial neural network (aNN) will be the geometric parameters of the selected transformer, which will be randomly generated within a range based on the system's electrical specifications. The outputs of the neural network will include the impedance matrix defining the transformer, the winding and core losses, and the average temperature reached throughout the transformer. The chosen core type is a U-core because it allows for easy stacking and custom area adjustment. AdFig. 1 Transformer geometry parametrization. ditionally, its open design enhances thermal dissipation, unlike the E-core type, where windings are concentrated on the central leg, limiting heat dissipation. The U-core also supports the use of parallel conductors due to its symmetrical structure, enabling windings to be distributed on each side of the core. The initial distribution of conductors will have the primary winding on one leg (blue conductors in Fig. 1), and the secondary winding will be placed on the opposite leg (yellow conductors in Fig. 1). Additionally, the artificial neural network (aNN) is constructed where both the primary and secondary windings are positioned on the same leg, while their parallel conductors are placed on the opposite leg. The exact electrical specifications of the converter are not yet determined, as the system is currently undergoing an optimization phase. In this application, the input voltage of the transformer may vary between 600 and 1200 V, with a current ensuring a power output of 150 kW, and an output voltage ranging from 200 to 700 V. The switching frequency will be between 40 and 60 kHz. Hence, the complete optimization process, including the neural network and preliminary calculations, is represented in Fig. 2. Fig. 2 Transformer design and optimization process based on analytical equations and aNN. Therefore, to ensure that the neural network provides results related to the specifications within the desired ranges, it will be trained using realistic random values. Assuming that a possible flux density 𝐵 could range between 50 and 150 mT, this determines ∫ 𝑉 𝑝 𝑑 𝑡 𝐵𝐴 𝑐𝑜𝑟𝑒 = 𝑁 𝑝 (1) Where the area 𝐴𝑐𝑜𝑟𝑒 could be described as 𝐶2 (see Fig. 1) and assuming square waveform for the voltage, it is possible to determine C as 𝑉 𝑝 𝐶 = √ 4 𝑓 𝑠𝑤 𝐵 𝑠𝑎𝑡 𝑁 𝑝 (2) Due to the large current flowing through the conductors, litz-wire conductors are the most suitable option. This type of conductor has an approximate packing factor 𝑘𝑙𝑤 of 0.5, meaning that only half of the conductor's volume consists of copper. Therefore, assuming a square litz-wire geometry (to simplify and accelerate the finite element simulation), the conductor width can be described as ∅ 𝑏𝑑 = √ 𝐴 𝑐𝑢 𝑘 𝐿𝑊 (3) Where the area of copper needed could be calculated as 𝐼 𝑅 𝑀 𝑆 𝐴 𝑐𝑢 = 𝐽 (4) Also, the self-inductance could be approximated to the magnetizing inductance as 𝑁 2 𝐿 = 𝑝 𝑚𝑎𝑔 𝑅 𝑒𝑞 (5) Furthermore, to prevent the generation of undesired geometries, H, W, and the gap are constrained to values relative to the overall geometry. This ensures that the resulting structures are both feasible and realistic. A total of 1000 distinct and random cases are generated based on these equations, which determine the geometry according to the electrical specifications. 3D simulations are then performed in Ansys Electronics Desktop using the Eddy Current solver to generate the impedance matrices and losses. Simultaneously, average temperatures in the model are generated using Icepak. 3 Design Guidelines for the Electromagnetic Optimization Once the neural network is trained, an iteration of cases is performed based on the final specifications. This results in a design space based on losses and temperature (see Fig. 3), as well as the core and winding volume, while also meeting the magnetizing and leakage inductances within a certain percentage. Subsequently, several cases are selected and manually optimized. To do so, two steps are followed: adjusting the magnetizing and leakage inductances if necessary, and optimizing the litz-wire conductor. 3.1 Inductance adjustment One of the obtained cases, shown in Fig. 4a, presents a primary winding with 6 turns and a secondary winding with 4 turns, both positioned on the same core leg, while their parallel windings are located on the other leg. This results in a magnetizing inductance of 27 μH and a leakage inductance of 5.7 μH, with winding losses of 180 W and core losses of 150 W. However, the magnetizing inductance should be 30 μH and the leakage inductance should be 3 μH, requiring a reduction in the leakage inductance. To achieve this, a partial interwinding is performed, where the secondary winding is inserted between the primary winding (see Fig. 4b). This reduces the leakage inductance to 3.3 μH and increases the magnetizing inductance to 30 μH. Additionally, winding losses are reduced by 6%. If further reduction of the magnetizing inductance is needed, the gap can always be adjusted, and if Balanced design zone Fig. 3 Artificial neural networks result where the power losses and the temperature are evaluated. 𝑐𝑡𝑒 2 𝑁 𝑠𝑡𝑟𝑎𝑛𝑑𝑠 𝐺 𝑝 ( 𝛾 ) 𝜇 𝑙𝑖𝑡𝑧 ′′ = 𝜎𝜔𝜇 𝐴 0 𝑏𝑢𝑛𝑑𝑙𝑒 (7) 2 ∞ μ ′′ ( 𝑥 ) 𝑥 μ ′ = ∫ 𝑙 𝑑 𝑥 + μ ′ 𝑙𝑖𝑡𝑧 π 0 𝑥 2 − ω 2 𝑐𝑡𝑒 (8) a further decrease in leakage inductance is required, a greater intrawinding should be considered. 3.2 Winding Optimization Due to the winding's diameter and the high number of strands, using brute-force simulations with Finite Element tools to optimize the litz-wire is neither efficient nor, in some cases, feasible. To overcome this limitation and perform the 3D eddy current analysis, the actual winding is replaced with a homogeneous conductor that preserves the same energy and losses. The new conductivity and complex permeability are described and calculated in [13] as N strands σ litz = 2A F ( γ ) R 𝑏𝑢𝑛𝑑𝑙𝑒 s DC (6) Where 𝛾 is defined by 𝑑/√2, 𝑑 is the diameter of each strand, δ is the skin depth of each strand, 𝑁𝑠𝑡𝑟𝑎𝑛𝑑𝑠 is the number of strands in the Litz-wire conductor, 𝑅𝐷𝐶 is the DC resistance per unit length of one strand, 𝐴𝑏𝑢𝑛𝑑𝑙𝑒 is the area of the total bundle that occupies the litz-wire conductor, μ′ is a value that ensures that the real permeability of the equivalent layer is 1 in low frequency and 𝛾 𝐹 𝑠 ( 𝛾 ) = 4 √ 2 𝑏𝑒𝑟0(𝛾)𝑏𝑒𝑖1(𝛾) − 𝑏𝑒𝑟0(𝛾)𝑏𝑒𝑟1(𝛾) [ 𝑏𝑒𝑟2(𝛾) + 𝑏𝑒𝑖2(𝛾) 1 1 𝑏𝑒𝑖 0 ( 𝛾 ) 𝑏𝑒𝑟 1 ( 𝛾 ) − 𝑏𝑒𝑖 0 ( 𝛾 ) 𝑏𝑒𝑖 1 ( 𝛾 ) − ] 𝑏𝑒𝑟 2 ( 𝛾 ) + 𝑏𝑒𝑖 2 ( 𝛾 ) 1 1 (9) 2 𝛾𝜋 𝐺 𝑝 ( 𝛾 ) = − √ 2 𝑏𝑒𝑟2(𝛾)𝑏𝑒𝑟1(𝛾) + 𝑏𝑒𝑟2(𝛾)𝑏𝑒𝑖1(𝛾) [ 𝑏𝑒𝑟2(𝛾) + 𝑏𝑒𝑖2(𝛾) 0 0 𝑏𝑒𝑖 2 ( 𝛾 ) 𝑏𝑒𝑖 1 ( 𝛾 ) − 𝑏𝑒𝑖 2 ( 𝛾 ) 𝑏𝑒𝑟 1 ( 𝛾 ) − ] 𝑏𝑒𝑟 2 ( 𝛾 ) + 𝑏𝑒𝑖 2 ( 𝛾 ) 0 0 (10) Where 𝑏𝑒𝑟𝑖 and 𝑏𝑒𝑖𝑖 are the real and imaginary part of the Kelvin functions of order i. Using this model, it is possible to test various litzwire configurations, including the number of strands and their diameter. According to the results provided by the aNN, the optimal current density in this case is approximately 3 A/mm². As explained in [14], once the total copper area is determined, it must be divided into strands to minimize the proximity effect. To mitigate the skin effect, the total copper area should be increased. Typical strand diameters are 0.1 mm, 0.07 mm, and 0.05 mm. If a 0.1 mm diameter is selected, 3.500 strands are required; for 0.07 mm, 5.000 strands are needed; and for 0.05 mm, 7.000 strands are necessary. By reducing the strand diameter from 0.1 mm to 0.07 mm reduces the proximity effect by 10%. However, further reducing it to 0.05 mm presents challenges, as the improvement is minimal (less than 2% compared to 0.07 mm), and ensuring Fig. 4 Proposed U-core transformer and winding distribution. proper tinning of all 7.500 strands is difficult. Additionally, the 0.05 mm strands are significantly more fragile than the 0.07 mm strands, increasing the risk of breakage and reducing the expected performance gains. If a 5.000 × 0.07 mm litz-wire is used and further reduction of skin effect losses is required, the number of strands (and thus the total copper area) should be increased. However, increasing to 7.500 strands only reduces losses by 5% while introducing the same issues associated with 0.05 mm strands. This is because, although conduction losses are reduced by 40%, proximity losses increase by 35%. (a) (b) Fig. 5 a) Schematic of how primary (in blue) passes through the secondary conductors (in yellow) to try to eliminate proximity losses. b) Real prototype built under these guidelines. (a) (b) Fig. 6 a) Smal signal transformer test. Secondary open. b) Smal signal transformer test. Secondary shorted. 3.3 Winding manufacturing guides The manufacturing process is crucial in transformer design to ensure the transformer is manufacturable. The proposed design uses ferrite blocks to form the transformer core, held together by metal brackets. Water-cooled heatsinks are integrated into the top and bottom parts of the transformer core. Winding interleaving is used to achieve the desired leakage inductance and eliminate the need for additional series inductance. This requires the primary winding to pass through the secondary, increasing losses. To minimize proximity losses, the strategy shown in Fig. 5 keeps the field generated by the primary winding connection and secondary winding orthogonal. 4 Experimental Results In this section, all the measurements and tests performed on the transformer are presented. Both small signal and large signal tests are described. 4.1 Small-Signal characterization Once the prototype is manufactured, magnetic characterization is performed to determine if the impedance meets the requirements or if the transformer needs to be adjusted. An impedance analyzer is used for this purpose. In Fig. 6 the real transformer is represented as a box with the primary terminals (P1, P2) and secondary terminals (S1, S2). The first measurement is done connecting the primary terminals to the impedance analyzer while secondary remains open. Thus, the sum of leakage and magnetizing inductance of the cantilever transformer model is obtained. terminal length. Therefore, the actual measurement when the secondary is shorted is 𝐿𝑙𝑘 + 𝑁2 ⋅ 𝐿𝑠ℎ𝑜𝑟𝑡 . In this case, the expected leakage inductance is higher than the parasitic 𝐿𝑠ℎ𝑜𝑟𝑡, which is in the nH range, and the turns ratio is small (1.5), so the measurement error in 𝐿𝑙𝑘 is negligible. The measured inductance matrix is presented. 𝐿 11 𝐿 12 Where = ( 34,2 𝜇𝐻 22,74𝜇𝐻) 22,74 𝜇𝐻 16,8𝜇𝐻 The second measurement is performed by shorting the secondary side. This test measures the leakage inductance of the transformer model. Care must be taken as the short is not ideal, and there will be some parasitic inductance due to the 4.2 Large-Signal characterization Testing the entire converter is not feasible with the available equipment, as dissipating 150 kW is challenging. Therefore, two separate tests are performed on the transformer to independently test the core and winding. 4.2.1 Open-Circuit test In the open circuit test, nominal voltage is applied to the primary side of the transformer while the secondary side is open. The voltage applied is a square waveform obtained using a SiC MOSFET full-bridge inverter. This causes the magnetic flux to range from -Bpk to Bpk, making the core undergo the entire hysteresis loop. Temperature is measured during the test until steady state is reached. Fig. 7 shows the maximum core temperature over time. 4.2.2 Short-Circuit test In the short circuit test, the secondary side is shorted. A capacitor is placed at the output of the two parallel SiC MOSFET half-bridge modules to resonate with the transformer's leakage inductance. The half-bridge configuration allows the modules to be put in parallel to withstand the current of 200 Arms. The voltage applied to the transformer is slowly increased until the nominal current is reached, and then the temperature is monitored until the steady state is achieved. Fig. 8a shows a thermal image of the winding temperature after 30 minutes of test duration, with a maximum temperature of 63.5°C. 𝐿 = ( ) 𝐿 12 𝐿 22 (11) k = 94.89% L mag = 30.8μH L lk = 3.4μH (12) Fig. 7 Max. core temperature at 130 mT excitations (Bpk) at 50 kHz. (a) (b) Fig. 8 a) Steadystate temperature of the winding at short-circuit test. b) Waveforms at shortcircuit test. 5 Conclusion In this paper, we presented a methodology that combines the use of an artificial neural network (ANN) based on finite element simulations and analytical equations to enable fast design iterations, making the design process more efficient. However, manual correction of the transformer design was necessary to fine-tune the magnetic parameters due to the series inductance integration requirement. To address this, design guidelines are provided based on an interwinding strategy to achieve the desired leakage and magnetizing inductance values. Additionally, litz-wire optimization is presented, along with construction techniques based on field orthogonality to reduce proximity problems. To validate the proposed methodology, a prototype was built, and various tests were conducted, evaluating its performance under both smalland large-signal conditions. Acknowledge Funded by the European Union under GA no 101101961 -HECATE. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or Clean Aviation Joint Undertaking. Neither the European Union nor the granting authority can be held responsible for them. The project is supported by the Clean Aviation Joint Undertaking and its Members. References [1] Sarlioglu, Bulent, and Casey T. Morris. "More Electric Aircraft: Review, Challenges, and Opportunities for Commercial Transport Aircraft." IEEE Transactions on Transportation Electrification, vol. 1, no. 1, 2015, pp. 54-64. doi:10.1109/TTE.2015.2426499. [2] Infographic / More Electric Aircraft, 2017. [Online]. Available: https://aertecsolutions.com/en/multimedia/infografia-el-avionmas-electrico/. [Accessed: 19-Jun-2024]. [3] 787 Electrical System, 2013. [Online]. Available: http://787updates.newairplane.com/787Electrical-Systems/787-electrical-system. [Accessed: 19-Jun-2024]. [4] Context and Objectives. Electric Aviation: FlightPath2050 and the Future of Flight, 2023. [Online]. Available: https://hecate-project.eu/about/context-and-objectives/. [Accessed: 18-Jun-2024]. [5] D. Izquierdo, G. Casas, A. G. Garriga, I. Castro, A. Zieminska-Stolarska, M. Sobulska, and I. Zbicinski, "Electrical Distribution System LCA for Future Regional Aircraft—Preliminary Definition of Methodology," Aerospace, vol. 10, no. 11, Art. no. 920, 2023. [Online]. Available: https://www.mdpi.com/2226-4310/10/11/920. doi: 10.3390/aerospace10110920. [6] U. Drofenik, "A 150kW Medium Frequency Transformer Optimized for Maximum Power Density," in Proc. 2012 7th International Conference on Integrated Power Electronics Systems (CIPS), 2012, pp. 1-6. [7] X. She, A. Q. Huang, and R. Burgos, "Review of Solid-State Transformer Technologies and Their Application in Power Distribution Systems," IEEE Journal of Emerging and Selected Topics in Power Electronics, vol. 1, no. 3, pp. 186-198, Sep. 2013. [8] Y. Wu, M. H. Mahmud, S. Christian, R. A. Fantino, R. A. Gomez, Y. Zhao, and J. C. Balda, “A 150-kw 99% efficient all-silicon-carbide triple-active-bridge converter for solar-plus-storage systems,” IEEE Journal of Emerging and Selected Topics in Power Electronics, vol. 10, no. 4, pp. 3496–3510, 2022 [9] Z. Guo, S. Rajendran, J. Tangudu, Y. Khakpour, S. Taylor, L. Xing, Y. Xu, X. Feng, and A. Q. Huang, “A novel high insulation 100 kw medium frequency transformer,” IEEE Transactions on Power Electronics, vol. 38, no. 1, pp. 112–117, 2023 [10] Z. Zhao, Y. Wu, F. Diao, N. Lin, X. Du, Y. Zhao, and G. Zhu, “Design and demonstration of a 100 kw high-frequency matrix core transformer for more electric aircraft power distribution,” IEEE Transactions on Transportation Electrification, vol. 8, no. 4, pp. 4279–4290, 2022. [11] R. Lu, J. Sheng, Y. Yao, C. Li, W. Li, and X. He, “Design and verification of a high ratio medium frequency transformer for subsea power conversion,” in 2021 IEEE 12th Energy Conversion Congress Exposition - Asia (ECCEAsia), 2021, pp. 886–892. [12] Z. Guo, C. Chen, R. Yu, and A. Q. Huang, “A high power 200kw medium frequency transformer with improved thermal management,” in 2023 IEEE Energy Conversion Congress and Exposition (ECCE), 2023, pp. 5587–5591. [13] A. Delgado, G. Salinas, J. A. Oliver, J. A. Cobos and J. Rodriguez-Moreno, "Equivalent Conductor Layer for Fast 3-D Finite Element Simulations of Inductive Power Transfer Coils," in IEEE Transactions on Power Electronics, vol. 35, no. 6, pp. 6221-6230, June 2020, doi: 10.1109/TPEL.2019.2949438. [14] A. Delgado, D. Schoenberger, J. Á. Oliver, P. Alou and J. A. Cobos, "Design Guidelines of Inductive Coils Using a Polymer Bonded Magnetic Composite for Inductive Power Transfer Systems in Electric Vehicles," in IEEE Transactions on Power Electronics, vol. 35, no. 8, pp. 7884-7893, Aug. 2020, doi: 10.1109/TPEL.2020.2965219.