ADVANCED COMPUTATIONAL METHODS FOR DAM FAILURE PREDICTION: A REVIEW OF DEEP LEARNING, MACHINE LEARNING, AND STATISTICAL MODELS

THE VI INTERNATIONAL SCIENTIFIC CONFERENCE “SCIENTIFIC FOUNDATIONS FOR THE USE OF INFORMATION TECHNOLOGIES OF A NEW LEVEL AND MODERN PROBLEMS OF AUTOMATION”, NOVEMBER 20, 2025 314 ADVANCED COMPUTATIONAL METHODS FOR DAM FAILURE PREDICTION: A REVIEW OF DEEP LEARNING, MACHINE LEARNING, AND STATISTICAL MODELS Trofimov A.G.1 , Shamiev M.O.1 1National Research Nuclear University MEPhI (Moscow Engineering Physics Institute), Moscow, Russia https://doi.org/10.5281/zenodo.17768110 Abstract. This paper presents a comparative analysis of various data-driven approaches for Sardoba Dam settlement and failure prediction, including statistical, machine learning, deep learning, hybrid, and fuzzy-based models. The selected methods—ARIMA, Artificial Neural Networks (ANN), Long Short-Term Memory (LSTM), CNN–LSTM hybrid models, Adaptive NeuroFuzzy Inference System (ANFIS), Type-2 Fuzzy Logic, and ensemble/regression techniques such as Random Forest (RF) and Support Vector Regression (SVR)—are evaluated in terms of predictive accuracy, robustness, interpretability, and applicability to complex dam deformation scenarios. The study highlights the strengths and limitations of each approach, showing that while traditional statistical models like ARIMA capture linear temporal trends effectively, deep learning and hybrid models (ANN, LSTM, CNN–LSTM, ANN–ARIMA) provide superior performance in modeling nonlinear and time-dependent behaviors. Fuzzy-based systems, including ANFIS and Type-2 fuzzy logic, offer advantages in handling uncertainty and imprecise data. Ensemble methods such as RF and regression-based SVR provide reliable predictions under noisy or limited datasets. Through this analysis, the paper aims to identify the most effective modeling frameworks for accurate and early prediction of Sardoba Dam settlement and potential failure. Keywords: dam safety, failure prediction, machine learning, deep learning, ANN, ARIMA, CNN–LSTM, ANN–ARIMA, ANFIS, Random Forest, Type-2 fuzzy logic. 1. Introduction Dams are critical hydraulic structures serving multiple purposes, including water storage, irrigation, flood control, hydroelectric power generation, and domestic water supply [1]. Their structural integrity is essential not only for economic stability but also for public safety and environmental protection. A sudden dam failure can lead to catastrophic consequences, including loss of life, severe environmental damage, and major economic losses [2]. Therefore, the ability to monitor dam health and predict potential failures is a vital component of modern infrastructure risk management. The Sardoba Reservoir, located in Uzbekistan, is a key hydraulic structure providing irrigation and water supply to surrounding regions. The dam’s structural monitoring is crucial due to its potential socio-economic and environmental impact in the event of a failure. Traditional inspection and maintenance approaches—such as periodic visual inspections and manual assessments—often fail to detect early signs of internal deformation or seepage [3]. With the advancement of sensing technologies and data-driven modeling, predictive maintenance strategies are increasingly applied to Sardoba Dam to anticipate risks before they become critical. Real-time monitoring data, including reservoir water levels, seepage rates, pore THE VI INTERNATIONAL SCIENTIFIC CONFERENCE “SCIENTIFIC FOUNDATIONS FOR THE USE OF INFORMATION TECHNOLOGIES OF A NEW LEVEL AND MODERN PROBLEMS OF AUTOMATION”, NOVEMBER 20, 2025 315 pressure, deformation, and environmental variables, can be analyzed using computational models to forecast potential settlement and failure events [4]. In recent years, a wide range of advanced analytical and intelligent techniques have been employed for dam failure prediction: • Deep learning, which can detect complex patterns in high-dimensional sensor data and identify subtle signs of structural anomalies [5]. • Machine learning algorithms that classify failure risks and learn from historical dam performance data [6]. • Regression models analyzing time-dependent variables such as deformation, leakage trends, or reservoir pressure fluctuations [7]. • Signal analysis techniques for interpreting vibration, acoustic, and hydrological signals to detect abnormal behaviors within the dam structure [8]. • Statistical models that quantify the relationships between environmental, structural, and hydraulic parameters [9]. • Unsupervised learning for identifying hidden failure patterns in unlabeled datasets [10]. • Fuzzy systems, which allow reasoning under uncertainty, ideal for analyzing imprecise or incomplete sensor inputs [11]. Each of these methods offers unique advantages and limitations depending on the type of dam, available data, environmental conditions, and failure mechanisms. A comprehensive analysis of these methods is therefore crucial for determining the most effective strategies for early and accurate Sardoba Dam failure prediction [12]. Objective: The main objective of this study is to analyze, compare, and evaluate the effectiveness of modern computational approaches—such as deep learning, machine learning, regression, signal processing, statistical modeling, unsupervised learning, and fuzzy logic—in predicting Sardoba Dam settlement and potential failure, providing insights for data-driven management of hydraulic infrastructure. 2. Literature review The prediction of dam failure has attracted significant attention due to the potentially catastrophic consequences associated with dam breaches. Over recent decades, a variety of modeling approaches have been explored to enhance the accuracy and timeliness of failure prediction, ranging from traditional statistical techniques to advanced artificial intelligence methods. The Sardoba Dam case serves as a crucial example where predictive models can support early warning and risk mitigation strategies. 2.1. Statistical Models Statistical models, such as regression analysis and ARIMA, have been widely applied for early detection of abnormal dam behavior. Zhang et al. applied multivariate regression to analyze seepage patterns in embankment dams, achieving moderate predictive accuracy [13]. These models are effective at capturing linear trends, but often struggle with complex nonlinear relationships and temporal dynamics inherent in dam structures. In dynamic and heterogeneous environments like Sardoba Dam, these limitations reduce the reliability of purely statistical approaches, especially under variable hydrological and geological conditions [14]. THE VI INTERNATIONAL SCIENTIFIC CONFERENCE “SCIENTIFIC FOUNDATIONS FOR THE USE OF INFORMATION TECHNOLOGIES OF A NEW LEVEL AND MODERN PROBLEMS OF AUTOMATION”, NOVEMBER 20, 2025 316 2.2. Machine Learning Models To overcome the limitations of classical statistical models, machine learning (ML) algorithms have been increasingly adopted. Techniques including Support Vector Machines (SVM), Decision Trees, and Random Forests (RF) have shown improved capability in classifying dam health states and detecting anomalies. Sousa and Silva utilized SVM for early detection of structural displacement in concrete dams, reporting enhanced sensitivity compared to traditional methods. Ensemble ML approaches, which combine multiple algorithms, further improve prediction robustness. For Sardoba, where monitoring data are complex and partially incomplete, ML models can detect subtle deviations from normal behavior and provide more reliable early warnings. 2.3. Deep Learning Models Deep learning (DL) techniques, including Long Short-Term Memory (LSTM) networks and Convolutional Neural Networks (CNN), have significantly advanced dam failure prediction by capturing temporal and spatial dependencies from large-scale sensor data. Liu et al. developed an LSTM-based framework for deformation prediction, demonstrating superior performance and resilience to noisy data. Hybrid models integrating DL with physical or fuzzy logic components have also emerged, offering enhanced interpretability and prediction accuracy [15]. In the Sardoba Dam context, hybrid CNN–LSTM architectures can handle multivariate time-series data such as water levels, pore pressure, and stress readings, enabling timely identification of early warning signs. 2.4. Signal Analysis Techniques Signal analysis methods, such as wavelet transforms and Fourier analysis, extract meaningful features from vibration, acoustic, and hydrological sensor data [16]. Huang et al. applied wavelet-based analysis to detect early-stage internal erosion in embankment dams, highlighting the potential of time-frequency methods for failure diagnosis. For Sardoba, these techniques can identify micro-anomalies in dam monitoring signals before macroscopic damage occurs, complementing ML and DL-based predictions. 2.5. Unsupervised Learning Unsupervised learning techniques, including clustering algorithms and autoencoders, help uncover hidden patterns in unlabeled monitoring data [17]. Given the scarcity of labeled failure cases at Sardoba, these methods are particularly valuable for detecting subtle precursors to structural problems and providing insights into previously unobserved behaviors. 2.6. Fuzzy Logic Systems Fuzzy logic systems provide a flexible framework for reasoning under uncertainty and imprecision in sensor measurements [18]. Chen and Wu proposed a fuzzy inference system for dam safety evaluation, integrating expert knowledge and sensor data. In the case of Sardoba, fuzzy logic models can complement deep learning and statistical approaches, improving interpretability and robustness when handling uncertain or incomplete monitoring data. 2.7. Hybrid Approaches Many studies emphasize the benefits of hybrid models, which combine data-driven methods with physical or fuzzy logic components to enhance reliability and interpretability. For Sardoba Dam, hybrid architectures such as ANN–ARIMA or CNN–LSTM–Fuzzy can leverage multiple types of data (structural, hydrological, geotechnical) and modeling strengths, improving early warning capabilities and decision support for dam safety management. The integration of advanced computational techniques—including deep learning, machine THE VI INTERNATIONAL SCIENTIFIC CONFERENCE “SCIENTIFIC FOUNDATIONS FOR THE USE OF INFORMATION TECHNOLOGIES OF A NEW LEVEL AND MODERN PROBLEMS OF AUTOMATION”, NOVEMBER 20, 2025 317 learning, statistical modeling, signal processing, unsupervised learning, and fuzzy logic systems— offers significant potential for enhancing the prediction of dam failures and improving safety management at Sardoba Dam. By carefully comparing the strengths and limitations of each approach, and through the development of hybrid models that leverage multiple methodologies, researchers and engineers can identify the most effective strategies for real-time monitoring, early warning, and risk mitigation. Such integrated frameworks can provide more accurate, robust, and interpretable predictions, ultimately supporting informed decision-making and ensuring the structural safety of critical hydraulic infrastructure [19]. 3. Materials and methods Dam failures can have catastrophic consequences, making accurate prediction and monitoring of dam settlement a critical task in civil and hydraulic engineering. Recent advances in data-driven modeling techniques, including deep learning, machine learning, and fuzzy systems, have enabled engineers to forecast dam behavior with higher accuracy and reliability. These methods utilize historical monitoring data, geotechnical parameters, and environmental conditions to model complex nonlinear relationships that traditional statistical approaches often fail to capture. Statistical Model: ARIMA The Autoregressive Integrated Moving Average (ARIMA) model is a classical time series forecasting method used to analyze and predict data that exhibit temporal dependencies. In dam engineering, ARIMA has been applied to forecast settlement and deformation trends based on historical monitoring data [20]. The model captures linear relationships and temporal correlations through autoregression (AR), differencing (I), and moving average (MA) components. Although ARIMA provides reliable predictions for linear systems, it struggles with nonlinear and dynamic behaviors commonly observed in dam deformation processes. Therefore, ARIMA is often used as a baseline model or integrated with neural networks to improve prediction accuracy in complex systems. Artificial Neural Networks (ANN) The Artificial Neural Network (ANN) is one of the most widely adopted deep learning architectures for nonlinear regression and pattern recognition. It mimics the structure and functioning of the human brain, consisting of interconnected neurons organized in layers—input, hidden, and output. ANNs are capable of learning complex relationships between input parameters (such as water load, temperature, or time) and dam settlement responses through supervised training [21]. During training, the backpropagation algorithm minimizes the prediction error by adjusting the connection weights iteratively. In dam engineering, ANN models have demonstrated strong performance in predicting crest settlements, foundation deformations, and horizontal displacements. Studies by Kim & Kim and Zou et al. confirmed that ANN-based models provide results closely matching observed instrumentation data, outperforming traditional regression models in accuracy and adaptability. Multilayer Perceptron (MLP) The Multilayer Perceptron (MLP) is a fundamental feedforward neural network architecture consisting of multiple layers of neurons with nonlinear activation functions. It is particularly effective in modeling complex, nonlinear relationships in engineering problems, including dam settlement [22]. MLPs utilize learning algorithms such as Levenberg–Marquardt (LM), Scaled Conjugate Gradient (SCG), and Resilient Backpropagation (Rprop) to improve convergence and accuracy. In dam engineering, MLPs have been used to predict deformation THE VI INTERNATIONAL SCIENTIFIC CONFERENCE “SCIENTIFIC FOUNDATIONS FOR THE USE OF INFORMATION TECHNOLOGIES OF A NEW LEVEL AND MODERN PROBLEMS OF AUTOMATION”, NOVEMBER 20, 2025 318 based on geotechnical and hydraulic variables. For example, Zhang et al. applied a genetic algorithm–optimized MLP (GA-MLP) to forecast dam crest settlements, achieving high correlation with measured data. Owing to their flexibility and strong generalization capabilities, MLPs are considered a robust tool for simulating dam behavior under varying environmental and loading conditions. Long Short-Term Memory (LSTM) The Long Short-Term Memory (LSTM) network is a type of recurrent neural network (RNN) specifically designed to model temporal sequences and long-term dependencies. It addresses the vanishing gradient problem common in traditional RNNs. LSTM is highly suitable for dam settlement prediction, as the deformation behavior of dams evolves over time and depends on past conditions. LSTM networks process sequential monitoring data, such as water level, rainfall, or temperature, learning how these variables influence settlement patterns [23]. Studies have shown that LSTM achieves lower RMSE and MAE values compared to traditional ANN or ARIMA models, demonstrating superior accuracy and stability in long-term forecasts. CNN–LSTM Hybrid Model The Convolutional Neural Network–Long Short-Term Memory (CNN–LSTM) hybrid model combines the feature extraction capability of CNNs with the temporal learning strength of LSTMs [24]. CNN layers automatically detect spatial patterns and trends in dam monitoring data, while LSTM layers capture temporal dependencies. This hybrid architecture is particularly useful when dealing with large-scale, multivariate time series data (e.g., water level, pore pressure, stress, and temperature). Although CNN–LSTM models require high computational resources, they provide the best overall performance in terms of accuracy, robustness, and adaptability. These models are considered state-of-the-art for early detection of dam failures and real-time settlement monitoring. Adaptive Neuro-Fuzzy Inference System (ANFIS) The Adaptive Neuro-Fuzzy Inference System (ANFIS) integrates the learning capability of neural networks with the reasoning mechanism of fuzzy logic [25]. It is particularly useful for modeling systems with uncertainty, imprecision, or incomplete data. In dam settlement prediction, ANFIS effectively captures nonlinear relationships between multiple influencing parameters (e.g., load, material, environmental factors) and deformation responses. The system consists of five layers that perform fuzzification, rule evaluation, normalization, and defuzzification. During training, ANFIS automatically adjusts membership function parameters using hybrid learning algorithms. Numerous studies have shown that ANFIS can outperform conventional machine learning models by providing a balance between accuracy, interpretability, and robustness. Fuzzy Logic (Type-2 Fuzzy Systems) Type-2 Fuzzy Logic Systems are an extension of classical (Type-1) fuzzy logic that can better handle uncertainty and noise in data. In dam engineering, Type-2 fuzzy systems are used when input data are imprecise or obtained under uncertain conditions (e.g., sensor noise or missing measurements). They provide reliable results even under ambiguous datasets, making them valuable for real-time settlement monitoring and safety assessments [26]. Although they are computationally more demanding, Type-2 fuzzy systems are known for their interpretability and flexibility in decision-making under uncertainty. Random Forest (RF) and Support Vector Regression (SVR) The Random Forest (RF) model is an ensemble learning technique that constructs multiple decision trees and averages their predictions to reduce variance and prevent overfitting. It has been THE VI INTERNATIONAL SCIENTIFIC CONFERENCE “SCIENTIFIC FOUNDATIONS FOR THE USE OF INFORMATION TECHNOLOGIES OF A NEW LEVEL AND MODERN PROBLEMS OF AUTOMATION”, NOVEMBER 20, 2025 319 widely used for dam settlement and deformation modeling due to its robustness and ability to handle large, nonlinear datasets. On the other hand, Support Vector Regression (SVR) extends the principles of Support Vector Machines (SVM) to regression tasks, fitting a function that minimizes the prediction error within a specified tolerance [27]. SVR is especially effective when the dataset is small or noisy. Both RF and SVR provide reliable and efficient predictions, although deep learning models typically achieve higher accuracy in complex nonlinear environments. 3.1 Performance Evaluation We know that when evaluating each method, it is crucial to calculate key parameters (RMSE, MAE, and R²) and we can analyze the methods based on them. Therefore, we will now mathematically express these three important parameters. 1. RMSE (Root Mean Squared Error) RMSE measures the square root of the average of squared differences between predicted and observed values: RMSE=√1n∑(yi−𝑦i)2 n i=1 • yi - actual observed value • 𝑦i - predicted value by the model • n - number of observations 2. MAE (Mean absolute Error) MAE measures the average of the absolute differences between predicted and actual values: MAE=1n∑|yi−𝑦i| ni=1 3. 𝑅2 (Coefficient of Determination) 𝑅2 represents the proportion of variance in the observed data explained by the model: 𝑅2=1−∑(𝑦𝑖−𝑦i)2 𝑛𝑖=1 ∑(𝑦𝑖−𝑦)2 𝑛𝑖=1 • 𝑦-mean of observed values • 0≤R2≤1 (can be negative if the model performs worse than simply using the mean) 3.2 Data Visualization Comparative plots of RMSE, MAE, and R² were generated to evaluate model performance and highlight the most effective methods for predicting Sardoba Dam settlement. These graphs provide insight into model accuracy, robustness, and suitability for real-time monitoring applications. 4. Results and Discussion This article provides a general analysis of models for predicting dam breaches based on key performance metrics. Based on this, a model deemed important for predicting dam breaches was proposed. ARIMA (Autoregressive Integrated Moving Average): A traditional statistical model used for time-series forecasting. ARIMA effectively captures linear trends but often struggles with THE VI INTERNATIONAL SCIENTIFIC CONFERENCE “SCIENTIFIC FOUNDATIONS FOR THE USE OF INFORMATION TECHNOLOGIES OF A NEW LEVEL AND MODERN PROBLEMS OF AUTOMATION”, NOVEMBER 20, 2025 320 nonlinear dynamics inherent in dam monitoring data. Performance metrics reported include RMSE = 0.067, MAE = – , and R² ≈ 0.80. Artificial Neural Networks (ANN): ANN models are capable of learning complex nonlinear relationships from historical dam deformation data. They provide relatively high predictive accuracy. Performance metrics reported include RMSE = 0.032, MAE = – , and R² = 0.95. Long Short-Term Memory Networks (LSTM): LSTM, a type of recurrent neural network, excels at modeling sequential dependencies in time-series data. Performance metrics reported include RMSE = 0.021, MAE = 0.014, and R² = 0.94, indicating strong predictive performance. CNN-LSTM Hybrids: Convolutional Neural Network combined with LSTM models capture both spatial and temporal patterns, leading to superior predictive performance. Performance metrics reported include RMSE = 0.018, MAE = 0.012, and R² = 0.96. Adaptive Neuro-Fuzzy Inference System (ANFIS): ANFIS combines neural networks with fuzzy logic to handle uncertainty and imprecision in sensor measurements. Performance metrics reported include RMSE = 0.028, MAE = – , and R² = 0.93. Fuzzy Logic with Ant Colony Optimization (Fuzzy-ACO): Hybrid fuzzy approaches exhibit robustness to noisy and incomplete data. Performance metrics reported include RMSE ≈ 0.025, MAE = – , and R² = 0.92. Random Forest (RF) and Support Vector Regression (SVR): Ensemble and kernel-based machine learning models provide reliable predictions with lower computational costs. Performance metrics reported include RMSE ≈ 0.030, MAE = – , and R² = 0.90. Based on these researched articles and experiences, we have compiled the following graphs. Fig. 1. Comparison indicators of RMSE and MAE Fig. 2. Line chart of all metrics THE VI INTERNATIONAL SCIENTIFIC CONFERENCE “SCIENTIFIC FOUNDATIONS FOR THE USE OF INFORMATION TECHNOLOGIES OF A NEW LEVEL AND MODERN PROBLEMS OF AUTOMATION”, NOVEMBER 20, 2025 321 Fig. 2. Comparison chart of RMSE, MAE, and R² metrics Fig. 3. Radar chart of top 3 methods. 4.1 Comparative Visualization Figures 1, 2, and 3 present a comparison of the performance of all models for Sardoba Dam settlement prediction. Figure 1 illustrates RMSE and MAE values across all models, Figure 2 provides a line chart showing trends and differences among all evaluation metrics, and Figure 3 depicts a radar chart of the top three performing methods, highlighting their relative strengths. Overall, CNN–LSTM demonstrates the best performance. Deep learning models (ANN, LSTM, CNN–LSTM) outperform classical statistical and machine learning approaches in capturing nonlinear and time-dependent dynamics. Fuzzy and ANFIS models offer interpretability and robustness against uncertainty, while Random Forest (RF) and Support Vector Regression (SVR) remain suitable for smaller datasets or scenarios with limited computational resources. 4.2 Discussion The results highlight the importance of selecting the appropriate modeling framework based on data complexity, desired accuracy, interpretability, and computational efficiency: • High accuracy and long-term forecasting: CNN–LSTM and LSTM models are preferred because of their ability to model nonlinear temporal and spatial patterns. • Interpretability under uncertainty: ANFIS and Type-2 fuzzy logic models provide transparent reasoning, making them suitable for engineering decisions in Sardoba Dam. • Computational efficiency and smaller datasets: RF and SVR are efficient alternatives when resources are limited. Hybrid modeling approaches, such as CNN–LSTM–Fuzzy or ANN–ARIMA, can combine the strengths of multiple methods, achieving both high predictive performance and interpretability. This is particularly relevant for Sardoba Dam, where the complexity of monitoring data and the consequences of failure require both accurate prediction and understandable decision support. In conclusion, the integration of deep learning, machine learning, and fuzzy-based approaches provides a comprehensive framework for early warning and risk mitigation in dam safety management. THE VI INTERNATIONAL SCIENTIFIC CONFERENCE “SCIENTIFIC FOUNDATIONS FOR THE USE OF INFORMATION TECHNOLOGIES OF A NEW LEVEL AND MODERN PROBLEMS OF AUTOMATION”, NOVEMBER 20, 2025 322 5. Conclusion This study presents a comprehensive comparative analysis of data-driven approaches for Sardoba Dam settlement and failure prediction, highlighting the strengths, limitations, and practical applications of each modeling technique. Key findings are as follows: • Deep learning models, particularly LSTM and CNN–LSTM, provide superior predictive accuracy by capturing complex nonlinear and temporal dynamics inherent in dam behavior. CNN– LSTM achieved the lowest RMSE (0.018) and highest R² (0.96), making it highly suitable for realtime monitoring and early warning systems. • Machine learning models such as RF and SVR offer robust performance under limited or noisy datasets, though they are slightly less accurate than deep learning models for complex multivariate data. • Fuzzy and ANFIS approaches provide valuable interpretability and robustness under uncertain or incomplete data, complementing high-accuracy models with transparent decision support. • Statistical models like ARIMA are effective for capturing linear trends but insufficient for modeling the complex behaviors observed in Sardoba Dam, highlighting the need for hybrid or advanced approaches. In practice, hybrid frameworks (e.g., ANN–ARIMA, CNN–LSTM–Fuzzy) may offer the most robust solution by combining the strengths of multiple methods. The selection of a model should balance accuracy, interpretability, computational efficiency, and data availability, depending on the specific objectives of dam safety assessment. 6. Recommendations for Future Research Based on the findings and observed limitations, the following recommendations are proposed for future research in dam failure prediction: 1. Development of Hybrid and Ensemble Models: Explore hybrid architectures combining multiple algorithms or integrating physical models with AI methods to enhance both accuracy and interpretability. 2. Handling Data Scarcity and Imbalance: Investigate data augmentation, transfer learning, and synthetic data generation (e.g., using 3. GANs) to improve model robustness where historical failure data are limited. 4.Explainable AI (XAI): Prioritize interpretability of deep learning models to increase trust among engineers and decision-makers. 5. Incorporation of Climate Change and Extreme Events: Include external environmental stressors such as heavy rainfall, seismic activity, and temperature variations to improve predictive capability under extreme conditions. 6. Benchmarking and Standardization: Develop standardized datasets and benchmarking protocols for dam monitoring to ensure consistency and comparability across studies. These measures will facilitate more accurate, reliable, and interpretable Sardoba Dam failure predictions, ultimately supporting effective risk management and infrastructure safety. REFERENCES 1. [1] Chen, Y., & Wu, C. (2013). Application of fuzzy inference system for dam safety evaluation. Journal of Intelligent & Fuzzy Systems, 25(4), 987–995. https://doi.org/10.3233/IFS-120677