Prediction Horizon Research Articles

The Cost of Regulating Effort: Reward and Difficulty Cues With Longer Prediction Horizons Have a Stronger Impact on Performance.

Many theories on cognitive effort start from the assumption that cognitive effort can be expended at will, and flexibly up- or down-regulated depending on expected task demand and rewards. However, while effort regulation has been investigated across a wide range of incentive conditions, few investigated the cost of effort regulation itself. Across four experiments, we studied the effects of reward expectancy and task difficulty on effort expenditure in a perceptual decision-making task (random-dot-motion) and a cognitive control task (colour-naming Stroop), and within each task comparted cues between short (cueing the next trial) and long (cueing the next six trials) prediction horizons. We found that participants used the cue information only when it was valid for multiple trials in a row. In the random-dot-motion task, a high reward expectancy resulted in better accuracy, especially in easy trials, but only with long prediction horizon. Similarly, in the Stroop task, the reward facilitation of reaction time was only observed after reward cues with a long prediction horizon. Together, our results indicate that people experience a cost to effort regulation, and that lower adjustment frequency can compensate for this cost.

Solar Radiation Prediction Using Decision Tree and Random Forest Models in Open-Source Software

The present research focuses on solar radiation prediction, which is important for energy production in thermal and solar systems. For this purpose, open-source software (Python) and a methodology involving the creation, implementation, and testing of specific machine learning models random forest (RF) and decision tree (DT) were used. The metrics used to identify the effectiveness of the models in predicting solar radiation were the coefficient (R2), the mean square error (MSE), and the mean absolute error (MAE). The evaluation of the two methods is presented in three cases: for one, two, and seven days. The results show that the RF model has better results than the DT, with MAE and MSE values of 36.96 and 4238.77, respectively, and a determination coefficient of 0.96. The study emphasizes the importance of selecting the appropriate model based on the prediction horizon to estimate solar availability and improve solar and thermal energy system planning.

Unified carbon emissions and market prices forecasts of the power grid

Carbon emissions and market prices forecasts of the power grid are of great importance for all electricity traders and consumers. Both forecasts enable flexible demand scheduling, ensuring sustainability and cost-efficiency. Many studies show the benefits of using both forecasts independently but not in combination, which remains an unexplored problem. The latest state-of-the-art techniques involve advanced statistical and machine learning algorithms leveraging seasonal patterns and exogenous forecasts. However, most of the reported studies deal only with problems of modeling and feature engineering, neglecting the forecast and model errors, which accumulate within the time-evolving power grid. This research aims to tackle these issues by introducing a versatile framework for short-term probabilistic forecasting of unified carbon emissions and market prices for electricity intra-day market participants. The approach utilizes the Hidden Markov Model for predictive estimation to determine the future energy mix of the country’s power grid. Furthermore, a novel optimization-based strategy, Moving Horizon Predictive Correction, is proposed to enhance the estimated energy mix performance, minimizing forecast and model errors. Subsequently, two separate recurrent neural networks are trained to provide probabilistic forecasts of carbon emissions and market prices, accounting for the stochastic dynamic of the power grid. A comparative analysis examines six case studies from various European countries and compares them with state-of-the-art forecasting methods. The results indicate that the proposed method can improve the qualitative measures by up to 53% for carbon emissions and up to 18% for market prices forecasts. Besides improving traditional point perditions, methods show significant increases in the quality of prediction intervals. Further application of the proposed forecasts is employed for a flexible 4–hour electricity consumption schedule. This showcases the usage of the proposed forecasts to find the best possible trade-off for low carbon emission, cost-effective electricity consumption time slots.

Temporal Attention-Enhanced Stacking Networks: Revolutionizing Multi-Step Bitcoin Forecasting

This study presents a novel methodology for multi-step Bitcoin (BTC) price prediction by combining advanced stacking-based architectures with temporal attention mechanisms. The proposed Temporal Attention-Enhanced Stacking Network (TAESN) integrates the complementary strengths of diverse machine learning algorithms while emphasizing critical temporal features, leading to substantial improvements in forecasting accuracy over traditional methods. Comprehensive experimentation and robust evaluation validate the superior performance of TAESN across various BTC prediction horizons. Additionally, the model not only demonstrates enhanced predictive accuracy but also offers interpretable insights into the temporal dynamics underlying cryptocurrency markets, contributing to both practical forecasting applications and theoretical understanding of market behavior.

Improved non‐cascaded continuous‐set model‐free predictive control scheme for PMSM drives

AbstractThis article presents an improved multiple‐step model‐free predictive control (MS‐MFPC) method for permanent magnet synchronous motor (PMSM) drives. Specifically, a reduced‐order MS‐MFPC based on an ultra‐local model, which employs only the input and output of the plant without addressing any motor parameters, is developed to enhance the robustness and reliability of the non‐cascaded PMSM system in the presence of parametric uncertainties. Meanwhile, the decoupled form of algebraic ultra‐local model is derived to extend the prediction horizon to multiple steps without increasing the computational burden. Besides, in order to improve the dynamic performance, an online adaptive modification approach is embedded into this suggested design to enhance the cost function calculation for the first time. Compared with the conventional continuous control set‐model predictive control (CCS‐MPC) method, our development not only achieves the smaller overshoot, settling time and steady‐state error, but also attenuates the inherent issue of system uncertainties in widely existing industrial application. Finally, the proposed MS‐MFPC methodology is experimentally assessed for a PMSM test bench, where steady‐state and transient‐state performance tests confirm the interest of the proposal.

An Ensemble Network for High-Accuracy and Long-Term Forecasting of Icing on Wind Turbines.

Freezing of wind turbines causes loss of wind-generated power. Forecasting or prediction of icing on wind turbine blades based on SCADA sensor data allows taking appropriate actions before icing occurs. This paper presents a newly developed deep learning network model named PCTG (Parallel CNN-TCN GRU) for the purpose of high-accuracy and long-term prediction of icing on wind turbine blades. This model combines three networks, the CNN, TCN, and GRU, in order to incorporate both the temporal aspect of SCADA time-series data as well as the dependencies of SCADA variables. The experimentations conducted by using this model and SCADA data from three wind turbines in a wind farm have generated average prediction accuracies of about 97% for prediction horizons of up to 2 days ahead. The developed model is shown to maintain at least 95% prediction accuracy for long prediction horizons of up to 22 days ahead. Furthermore, for another wind farm SCADA dataset, it is shown that the developed PCTG model achieves over 99% icing prediction accuracy 10 days ahead.

Enhanced forecasting of emergency department patient arrivals using feature engineering approach and machine learning

BackgroundEmergency department (ED) overcrowding is an important problem in many countries. Accurate predictions of ED patient arrivals can help management to better allocate staff and medical resources. In this study, we investigate the use of calendar and meteorological predictors, as well as feature-engineered variables, to predict daily patient arrivals using datasets from eleven different EDs across three countries.MethodsSix machine learning (ML) algorithms were tested on forecasting horizons of 7 and 45 days. Three of them – Light Gradient Boosting Machine (LightGBM), Support Vector Machine with Radial Basis Function (SVM-RBF), and Neural Network Autoregression (NNAR) – were never before reported for predicting ED patient arrivals. Algorithms’ hyperparameters were tuned through a grid-search with cross-validation. Prediction performance was assessed using fivefold cross-validation and four performance metrics.ResultsThe eXtreme Gradient Boosting (XGBoost) was the best-performing model on both prediction horizons, also outperforming results reported in past studies on ED arrival prediction. XGBoost and NNAR achieved the best performance in nine out of the eleven analyzed datasets, with MAPE values ranging from 5.03% to 14.1%. Feature engineering (FE) improved the performance of the ML algorithms.ConclusionAccuracy in predicting ED arrivals, achieved through the FE approach, is key for managing human and material resources, as well as reducing patient waiting times and lengths of stay.

Transformer-Based Model for Predicting Trajectories in Autonomous Vehicle-Pedestrian Conflicts: A Proactive Approach to Road Safety

Accurate prediction of pedestrian trajectories is crucial for conflict detection in autonomous vehicle (AV) operations. Existing models like Constant Velocity and Long Short-Term Memory (LSTM) networks have limitations. This paper introduces a Transformer-based model for predicting AV and pedestrian trajectories in urban conflict scenarios. Using attention mechanisms, the model dynamically weights features to improve predictions. Trained on the nuPlan dataset with over 1,500 hours of driving data, the Transformer model outperformed Constant Velocity and LSTM models across prediction horizons at three seconds. In Boston, it achieved an Average Displacement Error (ADE) of 2.071 meters for AV prediction, compared to 5.892 meters for Constant Velocity and 6.769 meters for LSTM. The model accurately predicts evasive actions; identifying 62.2% of AVs would accelerate and 41.6% would swerve in response to pedestrians in Boston. In Singapore, Transformers accurately predicted that pedestrians often swerve and accelerate, causing AVs to slow down without altering course.

Autonomous Vehicle Motion Control Considering Path Preview with Adaptive Tire Cornering Stiffness Under High-Speed Conditions

The field of autonomous vehicle technology has experienced remarkable growth. A pivotal trend in this development is the enhancement of tracking performance and stability under high-speed conditions. Model predictive control (MPC), as a prevalent motion control method, necessitates an extended prediction horizon as vehicle speed increases and will lead to heightened online computational demands. To address this, a path preview strategy is integrated into the MPC framework that temporarily freezes the vehicle state within the prediction horizon. This approach assumes that the vehicle state will remain consistent for a specified preview distance and duration, effectively extending the prediction horizon for the MPC controller. In addition, a stability controller is designed to maintain handling stability under high-speed conditions, in which a square-root cubature Kalman filter (SRCKF) estimator is employed to predict tire forces to facilitate the cornering stiffness estimation of vehicle tires. The double lane change maneuver under high-speed conditions is conducted through the Carsim/Simulink co-simulation. The outcomes demonstrate that the SRCKF estimator could provide a reasonably accurate estimation of lateral tire forces throughout the whole traveling process and facilitates the stability controller to guarantee the handling stability. On the premise of ensuring handling stability, integrating the preview strategy could nearly double the prediction horizon for MPC, resulting in the limited increase of online computation burden brought while maintaining path tracking accuracy.

Predicting freeway non-recurring congestion via a spatio-temporal deep learning approach

Freeway unexpected events, such as car crashes, result in non-recurring congestion, which does not follow repetitive patterns in space and/or time and increases the challenge of freeway management. Existing prediction approaches of Freeway Non-recurring Congestion (FNC) are either qualitative with information loss or single-step quantitative for a short period. To address these limitations, this paper develops a quantitative, interpretability-enhanced, and spatial-temporal approach using a 2-Encoder-Decoder Model with an Attention Layer (Att-2ED) for FNC prediction. The model leverages Long Short-Term Memory (LSTM) and Multilayer Perceptron (MLP) as two encoders to process time-series speed profiles and static crash features independently. It also employs an LSTM-based decoder to generate a spatio-temporal sequence output and an attention mechanism to prioritise the weights of input sequences at each time step. The model has been trained and tested using real-world freeway data. The proposed model demonstrates superior performance in predicting traffic conditions across both spatial and temporal dimensions compared to existing benchmarks, achieving a Mean Absolute Error (MAE) of 3.74 for 15-minute-ahead predictions. Its effectiveness is further underscored through a comparative analysis of prediction accuracy across various congestion levels and crash severities. For instance, the MAE ranges from 2.77 to 7.00 for severe crashes and from 2.68 to 6.47 for peak-hour crashes, with prediction horizons spanning from 5 min to 60 min ahead. Moreover, the attention mechanisms incorporated in the model provide interpretability-enhanced predictions, highlighting the significance of the last 10-min input in the prediction.

Dynamic survival prediction of end-stage kidney disease using random survival forests for competing risk analysis.

A static predictive model relying solely on baseline clinicopathological data cannot capture the heterogeneity in predictor trajectories observed in the progression of chronic kidney disease (CKD). To address this, we developed and validated a dynamic survival prediction model using longitudinal clinicopathological data to predict end-stage kidney disease (ESKD), with death as a competing risk. We trained a sequence of random survival forests using a landmarking approach and optimized the model with a pre-specified prediction horizon of 5 years. The predicted cumulative incidence function (CIF) values were used to generate a personalized dynamic prediction plot. The model was developed using baseline demographics and 13 longitudinal clinicopathological variables from 4,950 patients. Variable importance analysis for ESKD and death informed the creation of a sequence of reduced models that utilized six key variables: age, serum albumin, bicarbonate, chloride, eGFR, and hemoglobin. The models demonstrated robust predictive performance, with a median concordance index of 84.84% for ESKD and 84.1% for death. The median integrated Brier scores were 0.03 for ESKD and 0.038 for death across all landmark times. External validation with 8,729 patients confirmed these results. We successfully developed and validated a dynamic survival prediction model using common longitudinal clinicopathological data. This model predicts ESKD with death as a competing risk and aims to assist clinicians in dialysis planning for patients with CKD.

Efficiency Maximization of Stand-Alone HRES Based on Tri-Level Economic Predictive Technique

Renewable energy has been widely used in grid-connected and standalone hybrid renewable energy systems. These systems require a hybrid energy storage system due to the unpredictable climate and the inequality between the produced energy and the consumed energy. In this paper, a tri-level optimization method is used to optimize the sizing and the energy management of a standalone HRES, simplify the proposed optimization problem, and speed up the convergence process. Horizon prediction and weighting factor strategies are combined with the tri-level technique to find the most appropriate quantity of each element in the project and find the best energy management strategy. The objective function of the proposed methodology aims to minimize the total cost and improve the efficiency of the whole system. The proposed method was investigated on a standalone PV-WT with battery-hydrogen storage in different scenarios. The simulation results from the Matlab toolbox show that the performance indicators (cost and efficiency) are affected by the combination of the weighting factor and the forecasting index. The total productivity was improved by more than 2.5% in some scenarios while the investment cost and the running cost were reduced by values of 49.3% and 28.6%, respectively.

Generalized Malmquist orthogonal functions based model predictive control

The computational burden could be an issue of traditional model predictive control (MPC) in real-time applications. There are different procedures to cope with this problem such as using explicit MPC or approximation of the control inputs using orthogonal-basis functions. This paper presents the design procedure and experimental validation of a generalized discrete-time Malmquist orthogonal functions-based model predictive control (MbMPC) whose usage can decrease calculation concerns. To the authors’ knowledge, this is the first time the Malmquist functions are introduced into the MPC design. The discrete-time orthogonal Malmquist functions properties are used within this approach for approximating the future control input increments with a smaller number of parameters than in traditional MPC approaches. It is shown that the satisfactory controller and closed-loop system performances can be achieved using smaller number of tuning parameters. The performance of the proposed MbMPC is shown and compared with discrete-time Laguerre orthogonal functions based MPC (LbMPC) applied to a DC motor servo system. The simulation and experimental results demonstrate better IAE, ITAE, ISE and ITSE controller performance indices as well as faster average calculation time in comparison with the LbMPC approach. Additionally, the advantage of using the proposed method is emphasized by showing how the length of the prediction horizon affects the computational burden in comparison to the traditional MPC approaches.

Model predictive control parameter optimization method for autonomous vehicles

Aiming at the problem of insufficient tracking accuracy and long computational time of model predictive control algorithm in trajectory tracking of autonomous vehicles, a model predictive control parameter optimization method based on a nondominated sorting whale optimization algorithm was proposed. In this method, the parameter optimization problem of the model predictive control algorithm was transformed into a multiobjective optimization problem, with the predictive horizon, control horizon, and sample time as optimized objectives and the sum of squared lateral trajectory errors and computational time as optimized variables. The multiobjective optimization problem was solved using the nondominated sort whale optimization algorithm, and then the Pareto optimal solution set was obtained. The best solution was determined using the expert scoring method, continuous order weighted average operator method, game theory combination weighting method, and technique for order preference by similarity to the ideal solution method. Finally, the phase plane method is used to analyze the vehicle handling stability. The experimental results show that compared with the initial method, the lateral trajectory errors of the proposed method are reduced by 66.09% on average, and the computational time is reduced by 54.68% on average. Therefore, the proposed method is considered to be an effective method for parameter optimization of model predictive control.

Discrete‐Time Integral Fast Terminal Sliding Mode Predictive Control for Dynamic Positioning Ships With Lumped Uncertainties and Input Constraints

ABSTRACTThis paper proposes a robust discrete‐time integral fast terminal sliding mode predictive control (DIFTSMPC) scheme for DP ship trajectory tracking under the consideration of unmodeled dynamics, environmental disturbance, and input constraints. An improved single closed‐loop control strategy is designed by adopting the earth‐fixed reference velocity as the virtual velocity. The improved nonlinear PID discrete‐time integral fast terminal sliding mode is designed to achieve faster convergence. The unknown nonlinear dynamics and the environmental disturbance are combined as a lumped uncertainty term and estimated using the one‐step delay observation method. Then the corrected sliding model state is optimized to track the reference sliding reaching law in the prediction horizon. The lumped uncertainty estimation is integrated into the sliding mode prediction to guarantee the robustness of the control scheme. The stability of the proposed scheme has been verified. Comparison results demonstrate the effectiveness and superiority in chattering suppressing and constraint handling of the proposed scheme.

An Orbital-attitude Coupled Control Framework for a Full-scale Flexible Electric Solar Wind Sail Spacecraft in Orbital Transformation Missions

In this work, a novel orbital-attitude coupled control framework is developed for the electric solar wind sail (E-sail) spacecraft, aiming to handle the orbital transformation and attitude maneuver simultaneously. In the framework, the desired attitude parameters and the slow-varying voltage component are provided by the orbital control strategy, while the current attitude parameters and the fast-varying voltage component are manipulated by the attitude control strategy to approach their desired values. The orbital-attitude coupled characteristics of the E-sail spacecraft, including the flexibility-induced coupling effect, are fully described by the referenced nodal coordinate formulation. Considering the input saturation conditions, the governing equation for the orbital control strategy is then derived, in which the in-plane and out-of-plane displacement and velocity errors are prescribed as the state variables to be eliminated. An integral sliding mode control (ISMC) scheme is proposed to improve the robustness against the unmeasurable disturbance term. A model predictive control (MPC) scheme is introduced to enhance the convergence efficiency, where a quadratic optimization is performed to plan the desired attitude parameters and voltage components within the prediction horizon. To evaluate the control performance in the orbital transformation and attitude maneuver missions on the displaced non-Keplerian orbit, a series of scenarios with complex initial conditions are simulated under different control schemes, including the ISMC-MPC compound scheme. The results show that the control strategy designed under the rigid-body assumptions may not be feasible for the flexible E-sail spacecraft, while the investigated control strategy realizes the accurate and efficient convergence of the orbital and attitude variables on both the rigid and flexible E-sail spacecraft with the tether deformation stabilized.

Evaluation of machine learning-based regression techniques for prediction of diabetes levels fluctuations.

Middle-Aged and Elderly people today face a variety of health problems as a result of their modern lifestyle, which includes increased work stress, less physical activity, and altered food habits. Because of Complications arising, diabetes has become one of the most frequent, severe, and fatal illnesses around the world. Therefore, inaccurate measurements of blood glucose levels can seriously damage vital organs. Several strategies for long-term glucose prediction have been proposed in the literature. Unfortunately, these methods require the patient to identify their daily activities, which can be error-prone, such as meal intake, insulin injection, and emotional aspects. This paper suggests using continuous glucose monitoring (CGM) of 14733 patients, with three assistance factors to predict blood glucose levels independently of other parameters, hence reducing the burden on the patients. To support this an Artificial Neural Network (ANN), Binary Decision Tree (BDT), Linear Regression (LR), Boosting Regression Tree Ensemble (BSTE), Linear Regression with Stochastic Gradient Descent (LRSGD), Stepwise (SW), Support Vector Machine (SVM), and Gaussian process regression (GPR) were investigated. The result indicated that The highest classification accuracy of (92.58%) has been achieved by BDT followed by BSTE (92.04%) and GPR (88.59%). The obtained average of root means square error (MSE) was 1.64, 1.67, 1.69, mg/dL for prediction horizon (PH) respectively to GPR, BSTE, and ANN.

Shortcomings in the Evaluation of Blood Glucose Forecasting.

Recent years have seen an increase in machine learning (ML)-based blood glucose (BG) forecasting models, with a growing emphasis on potential application to hybrid or closed-loop predictive glucose controllers. However, current approaches focus on evaluating the accuracy of these models using benchmark data generated under the behavior policy, which may differ significantly from the data the model may encounter in a control setting. This study challenges the efficacy of such evaluation approaches, demonstrating that they can fail to accurately capture an ML-based model's true performance in closed-loop control settings. Forecast error measured using current evaluation approaches was compared to the control performance of two forecasters-a machine learning-based model (LSTM) and a rule-based model (Loop)-in silico when the forecasters were utilized with a model-based controller in a hybrid closed-loop setting. Under current evaluation standards, LSTM achieves a significantly lower (better) forecast error than Loop with a root mean squared error (RMSE) of 11.57 ±0.05 mg/dL vs. 18.46 ±0.07 mg/dL at the 30-minute prediction horizon. Yet in a control setting, LSTM led to significantly worse control performance with only 77.14% (IQR 66.57-84.03) time-in-range compared to 86.20% (IQR 78.28-91.21) for Loop. Prevailing evaluation methods can fail to accurately capture the forecaster's performance when utilized in closed-loop settings. Our findings underscore the limitations of current evaluation standards and the need for alternative evaluation metrics and training strategies when developing BG forecasters for closed-loop control systems.

PHASE TRANSITION-BASED OPTIMIZED ADAPTING MODEL FOR REAL-TIME SEIZURE PREDICTION APPLICATION

Epilepsy is a dynamic process that will undoubtedly continue throughout an individual’s lifetime. One part of the brain may get affected, other brain regions also can possibly get involved in this disorder. Such seizures are sometimes life threatening; therefore, a warning signal is required to provide information to their respective caretakers in order to ensure the safety of the epileptic patients. The following are the four states of seizure: interictal (between seizures), preictal (before seizure), ictal (seizure), and postictal (after seizures). The ability to accurately distinguish the unique preictal state from the other seizure states is essential for seizure prediction. The Seizures can only be predicted by early detection of the preictal condition, which helps to differentiate seizure prediction from seizure detection. Preictal signals, which typically appear minutes to hours before a seizure occurs, are predicted in order to prevent seizures. In this paper, the discussions on the seizure predictions and its procedures are done. An optimized adapting model based on the phase transitions for real-time seizure prediction is developed and implemented in this study. We proposed a classification model called as enhanced convolutional neural networks (ECNNs) that has been tuned for speeding convergence and reducing model complexity utilizing the Fletcher Reeves Algorithm (WO-FRA) based on Walrus Optimization. Furthermore, the Phase Transition Predictor (PTP) based on the Kullback–Leibler (KL) divergence yields a Premium Seizure Prediction Horizon (PSPH). The suggested model’s empirical results, which were verified using 1000 EEG recordings from the CHB-MIT, NINC, and SRM databases, outperform the current techniques with 98% accuracy, 0.07 h[Formula: see text] false prediction rate and with 99% sensitivity.

Groundwater level predictions in the Thames Basin, London over extended horizons using Transformers and advanced machine learning models

This study breaks new ground by using the Temporal Fusion Transformer (TFT) method for groundwater level prediction, addressing the complex dynamics of the Thames Basin aquifer in England. Our research combines extensive hydrological data collected from the Thames Basin with advanced machine learning, where a complex network of rivers and streams substantially affects groundwater dynamics. Unlike previous studies, this research focuses on long-term forecasting with deep learning, offering, for the first time, a 60-day prediction horizon based on daily data. To rigorously examine the model performance and robustness on new, unseen data, we applied the walk-forward validation method and other matrices such as RMSE and R2 coupled with the Holdout technique. The models used were Long Short-Term Memory (LSTM), Attention-based LSTM, LSTM with Bayesian optimisation, Attention-based LSTM with Bayesian optimisation and TFT. They were used on the basin's Chalk, Jurassic Limestone, and Lower greensand aquifers. Whilst both LSTM models were optimised using the Bayesian technique, TFT was applied for its inherent capability in complex time series. Our methodology processed historical groundwater and rainfall data from 2001 to 2023, accounting for the potential lag in aquifer response to the proximity of the river system. The dataset served as training, validation, and holdout for each model, focusing on capturing the dynamic temporal fluctuation. The results clearly showed the superiority of the TFT model in all aquifer types compared to other models across all horizons. The Limestone had the greatest result in the 7-day projections, with an RMSE of 0.02 and R2 of 0.98; Whilst the Chalk and Lower greensand, had RMSEs of 0.03 with R2 values of 0.75 and 0.95, respectively. The Limestone aquifer performed best for the 30-day horizon again (RMSE = 0.06, R2 = 0.85), with the Chalk and Lower greensand aquifer yielding RMSE of 0.04 and 0.12 and R2 values of 0.64 and 0.74, respectively. In the 60 days predictions, the best results were observed in the limestone aquifer with RMSE of 0.09 and R2 of 0.65 in holdout validation. However, in chalk and lower greensand aquifers, the TFT showed RMSEs of 0.05 and 0.15 and R2s of 0.45 and 0.58, respectively. Traditional LSTM models demonstrated limited predictive power compared to the main model TFT, while the attention mechanism slightly improved the accuracy. This study not only sets a new benchmark in hydrological modelling accuracy but also highlights the potential of advanced machine learning in managing complex aquifers and predicting the water table.