Sep 2026· Journal of Transportation Engineering Part A Systems· 0 citations· 37 references
TL;DR
A novel spatiotemporal Kalman filter network (ST-KFNet) framework for metro demand forecasting by integrating an autoregressive integrated moving average module, a Kalman filter (KF) module, and a convolutional neural network (CNN)-based variational autoencoder (VAE) module is proposed.
Abstract
Accurate passenger demand prediction is crucial for efficient metro operations, yet remains challenging due to complex network structures, temporal variability, and travel uncertainties. This paper proposes a novel spatiotemporal Kalman filter network (ST-KFNet) framework for metro demand forecasting by integrating an autoregressive integrated moving average (ARIMA) module, a Kalman filter (KF) module, and a convolutional neural network (CNN)-based variational autoencoder (VAE) module. ARIMA constructs the state prediction equation, further enhanced by a neural-network-based predictor. CNN-based VAE captures spatial dependencies and temporal dynamics, refining the KF’s state transition for improved accuracy. The KF then provides real-time, online predictions, effectively handling both stationary and nonstationary uncertainties. By combining the strengths of statistical modeling and deep learning, ST-KFNet achieves high predictive accuracy under normal conditions and robust performance under uncertainty, such as unplanned station closures. Empirical evaluations on a metro data set from Hangzhou show that ST-KFNet consistently outperforms existing benchmarks, including graph convolutional network (GCN)-based methods as well as ARIMA, KF, and CNN-based VAE models.
Accurate short-term metro origin–destination (OD) demand forecasting is essential for intelligent passenger flow management and urban rail transit operation. However, forecasting large-scale metro OD demand remains challenging due to its high dimensionality, nonlinear spatiotemporal dependencies, and demand uncertainty. To address these challenges, this paper proposes a Tensor Network Extended Kalman Filter (TNEKF) framework for short-term metro OD demand forecasting. First, metro OD demand is formulated as a nonlinear dynamic state-space prediction problem, where a multi-input multi-output Volterra model is adopted to characterize the nonlinear relationship between historical passenger demand and future OD flows. An Extended Kalman Filter (EKF) is then developed to recursively estimate the latent model parameters and continuously refine demand prediction using newly available observations. To improve computational efficiency for high-dimensional OD systems, both the latent state vector and covariance matrix are represented using low-rank tensor network structures, and all recursive filtering operations are implemented through tensor-network contractions without explicitly constructing large-scale matrices. Experiments on real-world smart-card data from the Hangzhou metro system demonstrate that the proposed method consistently outperforms ARIMA, conventional EKF, and several state-of-the-art spatiotemporal prediction models in terms of MAE, RMSE, and MAPE. Compared with the best-performing baseline of the whole-day scenario, the proposed method reduces MAE, RMSE, and MAPE by 30.2%, 9.8%, and 6.3%, respectively. Furthermore, the proposed framework exhibits strong robustness under disruption scenarios, demonstrating its effectiveness and scalability for large-scale metro OD demand forecasting.
Aijing Su, Bing Wu, Xiaoxing Fang· ISPRS International Journal...· 0 citations
Accurate wind vector prediction is essential for renewable energy utilization and power system stability, yet existing methods struggle to jointly model local dynamics, global structures, and temporal robustness. To address this limitation, an Adaptive Multi-Branch Heterogeneous Fusion Wind Prediction Network (AMBHFN) is proposed. Local dynamic, global structural, and temporal robustness modeling are assigned to dedicated heterogeneous branches, whose outputs are coordinated through the Adaptive Multi-Branch Prediction Collaboration Mechanism (AMBPC). Multi-source meteorological variables and terrain information are used for local dynamic modeling, while global spatiotemporal structures are captured by a 3D U-shaped fully convolutional branch and temporal robustness is enhanced by an iTransformer-based multi-agent branch with graph convolution. Experiments on ERA5 data show that AMBHFN outperforms eight retrained baselines over the 0–23 h forecast horizon, with an average error reduction of more than 12%. At the first forecast step, the root mean square error (RMSE) and mean absolute error (MAE) are 0.33 m/s and 0.25 m/s, respectively. Under the strict 22.5° threshold, wind direction forecast accuracy (WDFA) reaches 97.72% at 0 h and 78.06% at 6 h. Fine-tuning in two target regions reduces the 13–23 h RMSE to 1.54 and 1.96. Statistical tests confirm significant improvements over MFWPN, and ablation studies verify the complementarity of the three branches. With 128 giga floating-point operations (GFLOPs) and a 22 ms per-sample forward inference time, AMBHFN achieves a competitive balance among accuracy, stability, and efficiency.
In smart mobility networks, accurate vehicular flow forecasting is of critical importance, enabling efficient, robust, user- and environment-friendly management of devices, technologies, and systems. However, current short-term traffic prediction algorithms frequently face challenges of computational inefficiency and limited predictive precision. To overcome these limitations, this paper introduces an LSTM model optimized through Bayesian optimizer (BO-LSTM) to enhance prediction accuracy. The traffic data is first preprocessed through data augmentation using random sampling and scaling of traffic counts, which is particularly suitable for traffic time series data as it preserves temporal patterns while increasing data diversity and robustness against demand fluctuations. After this augmentation step, the data is standardized to bring the input features to a similar range. The LSTM model is trained using Bayesian optimization for hyperparameters tuning, including the learning rate, dense layers, number of iterations, and dropout rate, within an acceptable range. The root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) are computed for three different datasets, measured in number of vehicles per minute, yielding respective values of 0.0442, 0.0353, and 2.91%, 0.0216, 0.0173, and 1.84%, and 0.4426, 0.3541, and 2.63%. These results correspond to improvements of 34.5, 59.5, and 32.6% in RMSE, respectively, when compared against Attention-LSTM and temporal convolutional network (TCN) models. This proves the accuracy of the proposed model. The developed scheme outperforms counterpart models, advocating its potential to enhance dynamic traffic management and intelligent signal coordination systems. Such capability enables more accurate short-term flow estimates, thereby reducing average vehicle waiting times and improving intersection-level signal responsiveness.
Hamza Murad Khan, Anwar Khan, Muhammad Imran Majid et al.· The Arabian journal for scie...· 0 citations
Accurate short-term traffic prediction is a critical component of intelligent transportation systems (ITS), yet it remains challenging due to nonlinear temporal dynamics, evolving spatial dependencies, and uncertainty in real-time urban traffic data. This paper proposes a novel uncertainty-aware deep ensemble spatiotemporal forecasting framework integrating Dynamic Graph Convolutional Networks (DGCN), Temporal Transformers, and CNN–LSTM hybrid models. A confidence-guided ensemble fusion strategy dynamically weights individual predictions using Bayesian uncertainty estimation. Experiments conducted on real-time Bhopal city traffic data demonstrate significant improvements over state-of-the-art baselines, achieving up to 90% performance gains during peak and abnormal traffic conditions.
Smart-building load forecasters are often trained offline on dense, multivariate, high-frequency data, but deployment may provide only hourly, feature-limited inputs. Missing features must then be reconstructed, and their errors can propagate through the model. If this input uncertainty is not reflected, prediction intervals may become miscalibrated, affecting demand-response scheduling. Our work examines where uncertainty should be placed once inference inputs are reconstructed. We develop a unified one-day-ahead probabilistic forecasting framework that aligns temporal resolution, reconstructs the unavailable inputs, and derives causal features, and we compare a modular post-hoc residual-quantile scheme with an integrated in-model quantile-learning scheme. The comparison uses three mid-scale Deep Learning (DL) backbones: recurrent, hybrid recurrent, and attention-based Temporal Fusion Transformer (TFT) models, under identical inputs, forecasting horizon, preprocessing rules, and training budgets. Results show that uncertainty placement is backbone-dependent. Integrated quantile learning is most reliable with the TFT, yielding 2.2-3.6% MAPE and 28-83W RMSE on the labeled test window, while producing intervals about 5x narrower than the modular intervals at the closest-to-nominal coverage level. Diebold-Mariano tests support the TFT ranking and the mixed behavior of the recurrent backbones. A reconstruction-sensitivity test shows that reconstructed inputs increase the Quantile Score (QS) by 106% while interval width remains nearly unchanged, indicating that the model does not automatically absorb reconstruction-induced uncertainty. Robustness checks against non-DL baselines and seasonal hold-out weeks support this ranking. Our results expose the limits of post-hoc residual quantiles when inference depends on reconstructed inputs.
S. Al-Shareeda, Gulcihan Ozdemir, H. Jeon· Electric power systems resea...· 0 citations
With the rapid growth of smart-building applications, accurate forecasting of building energy consump- tion faces two critical challenges: some operational features (e.g., occupancy or equipment load) cannot be predicted accurately or may be entirely unavailable due to sensor malfunctions at prediction time, and com- plex nonlinear, long-range temporal dependencies are difficult to capture with traditional machine-learning models. Conventional approaches, such as Seasonal AutoRegressive Integrated Moving Average with eXoge- nous regressors (SARIMAX), Random Forest (RF), eXtreme Gradient Boosting (XGBoost), and Support Vector Machine (SVM), require extensive feature engineering, often assume linearity or stationarity, and show reduced accuracy when only sparse inputs are provided. To overcome these limitations, we introduce a Knowledge Distillation (KD) framework in which a GPT-style Transformer-based Building Energy Forecast- ing (GPT-BEF) teacher model is first trained on rich historical multivariate time series to capture complex temporal dependencies in building energy use. Its knowledge is then transferred into a reduced-feature GPT- BEF+KD student model that uses only two temperature features. The resulting GPT-BEF+KD model achieves approximately 25% lower Root Mean Squared Error (RMSE) than a two-feature RF baseline, sug- gesting its potential for lightweight forecasting under limited sensor availability. The analysis demonstrates that the student model effectively inherits part of the teacher’s ability to balance long-term and short-term temporal signals. By combining archive-driven full-feature learning with feature-constrained inference, this work provides an initial demonstration of GPT-style knowledge distillation for sparse-feature building en- ergy prediction and lays the foundation for future advances in multimodal fusion, closed-loop control, and uncertainty-aware energy management.
Yusen Wang, Xilei Dai, Zhenhong Lin et al.· Energy Use· 0 citations
Related blog posts
MIT News · Artificial Intelligence· news.mit.eduJul 15, 2026
Assistant Professor Pat Pataranutaporn describes a new interface that lets everyday users glimpse inside an AI's neural network before their chatbot ever says a word.
MIT News · Artificial Intelligence· news.mit.eduJul 6, 2026
PhD student Rachel Sava, winner of the Envisioning the Future of Computing Prize, explores transformative improvements and dystopian risks of neural technology.