Aug 2026· Advanced Electromagnetics· Vol 15, pp. 5064-5074· 0 citations
TL;DR
A hybrid spatiotemporal prediction framework integrating an improved Long Short-Term Memory (LSTM) network with dynamic graph embedding for deep feature mining and coordinated forecasting and offers methodological support for intelligent electromagnetic energy management and resilient smart power systems is proposed.
Abstract
Accurate prediction of distributed photovoltaic (PV) output is essential for modern smart grids and electromagnetic energy infrastructure, where renewable generation exhibits strong nonlinearity and long-term temporal dependence due to cloud occlusion and meteorological variations. Traditional forecasting methods often struggle to characterize cross-node interactions and accumulate prediction errors under complex operating conditions. To address these issues, this paper proposes a hybrid spatiotemporal prediction framework integrating an improved Long Short-Term Memory (LSTM) network with dynamic graph embedding for deep feature mining and coordinated forecasting. A multi-layer residual LSTM with adaptive attention first models long-term dependencies and transient fluctuations from preprocessed time series data. A dynamic graph is then constructed according to feeder connectivity and geographical proximity, where Dynamic GraphSAGE generates node embeddings to capture evolving spatial relationships. Temporal features and graph representations are fused through a Graph Attention Network (GAT) and multi-layer LSTM to jointly model spatiotemporal interactions, while TimeGAN-based sparse data completion and Bayesian optimization further enhance robustness and parameter adaptation. Experimental results demonstrate RMSE values of 0.10, 0.12, and 0.13 for 30 min, 3 h, and 6 h forecasting horizons, respectively, with training speed 20% faster than GCN-LSTM, only a 20% RMSE increase under σ = 0.1 noise, and RMSE remaining below 0.14 at unseen sites. The proposed framework provides an effective solution for distributed PV forecasting and offers methodological support for intelligent electromagnetic energy management and resilient smart power systems.
Experimental results on various publicly available PV power datasets show that BiMS-DGCN consistently outperforms several state-of-the-art forecasting methods, thereby improving the accuracy of multi-site photovoltaic power forecasting.
Xin-Qi Liu, Peng Li, Linong Du et al.· International journal of pat...· 0 citations
Short-term load forecasting for modern smart grids must jointly model long-range temporal patterns, correlations across many metering points, and the uncertainty required for operational decision-making. We propose ST-DGT-UA, a spatio-temporal dynamic graph Transformer that (i) embeds heterogeneous covariates with convolutional projections and Time2Vec, (ii) learns a time-varying adjacency matrix from node representations, (iii) couples spatial graph attention with temporal ProbSparse attention for efficient long-sequence modeling, and (iv) outputs multiple conditional quantiles trained with multi-quantile Pinball loss. Experiments on GEFCom2014 and the UCI Electricity Load Diagrams datasets show that ST-DGT-UA achieves MAPE/RMSE of 1.92/145.6 on GEFCom2014 and 2.65/42.3 on UCI, and improves probabilistic quality with Pinball Loss 0.024 and CRPS 0.043. These results indicate that learning dynamic, data-drivenspatial dependencies is beneficial for multi-node load forecasting where correlations evolve over time.
Accurate photovoltaic (PV) power forecasting is crucial for reliable grid operation, energy management, and the effective integration of PV systems into modern power grids. In this study, a short-term PV power forecasting framework is designed based on a residual multi-scale network enhanced with a modified coordinate attention module (MCAM). The active PV power is decomposed into intrinsic mode functions (IMFs) using complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), and the obtained components are combined with radiation, temperature, and wind speed variables. The feature matrix is transformed into a multi-channel tensor through a sliding-window-based representation, enabling the network to learn feature–time dependencies in a structured form. The proposed architecture includes three residual multi-scale stages with cascaded convolutional branches to extract local and contextual patterns from different kernels. The MCAM is integrated after each multi-scale block to emphasize the spatio–temporal features while suppressing less relevant features. Experiments are conducted on real-world PV data from the Yulara solar power system for the 1-h, 2-h, and 3-h forecasting horizons. The proposed model achieves R2 values of 0.9950, 0.9902, and 0.9836, with RMSE values of 0.4400, 0.6185, and 0.7994 for 1-h, 2-h, and 3-h horizons, respectively. The effectiveness of the proposed method is validated through comprehensive comparisons with benchmark models and evaluations under seasonal and weather conditions.
A lightweight dynamic spatio-temporal graph neural network, EdgeLite-DSTGNN, is proposed to address the intensified spatio-temporal coupling of line-loss rates, complex state correlations, and limited edge-terminal resources in distribution networks with high renewable-energy penetration. The method treats branches as graph nodes and constructs a dynamic sparse graph by integrating physical topology, time-varying electrical relationships, and adaptive associations. Shared features are extracted using depthwise separable dilated causal convolutions, Top-k sparse graph attention, and gated fusion to jointly perform multi-step line-loss-rate prediction and identify high line loss, overload, voltage violations, and reverse power flow. Knowledge distillation, structured pruning, INT8 quantization, and event-triggered graph construction are further incorporated for edge deployment. Experiments on modified IEEE 33- and 69-bus systems show that, compared with PC-GAT, the proposed method reduces MAE, RMSE, and MAPE by 13.76%, 12.80%, and 13.58%, respectively, while improving Macro-F1 by 1.7 percentage points. The final INT8 model reduces file size, peak memory, inference latency, and energy consumption by 93.8%, 73.0%, 81.8%, and 81.7%, respectively, and the event-triggered mechanism lowers the average end-to-end latency from 12.8 ms to 9.3 ms.
Hai-Yan Wang, Ye Yuan, Xin-Ping Yuan et al.· International journal of pat...· 0 citations
Predicting the remaining useful life (RUL) is essential for effective predictive maintenance. Spatio-Temporal Graph Neural Networks (ST-GNNs), which can model both temporal and spatial relationships by representing time series data as a sequence of graphs, have shown exceptional performance in RUL prediction. However, current ST-GNNs face several drawbacks. First, they require domain expertise or significant computational power to establish graph structures prior to deploying GNNs. Second, the models are restricted to capture temporal dependencies within a predefined fixed-size lookback window. This restriction ignores the common issue of varying time series lengths, leading the prediction model to miss short-term or long-term dependencies. Finally, conventional models often fail to capture the inherent relationships between samples generated from adjacent time windows, which are crucial for improving both the accuracy and robustness of predictions. To address the aforementioned issues, we introduce a novel framework called Multi-Term Fourier Graph Neural Network with Sample Relationship Learning (MTFGN-SRL). Rather than treating the sample as a sequence of graphs, we consider it as a single complete graph and utilize a Fourier Graph Neural Network (FGN) to capture the spatio-temporal information in the frequency domain. We propose a multi-term learning module that utilizes multiple lookback windows to generate samples with varying terms, which are then fed into the FGN to enhance the extraction of useful information from the data. Finally, we develop a sample relationship learning module by training a heterogeneous GNN to identify inter-sample relationships, resulting in enhanced accuracy and robustness in predictions. Evaluations on the CMAPSS dataset demonstrate MTFGN-SRL's superior performance over state-of-the-art methods in RUL prediction.
Ya Song, Laurens Bliek, Yao-Xin Wu et al.· 0 citations
A closed-loop integration framework that combines a particle swarm optimization (PSO)-based Temporal Fusion Transformer-Graph Attention Network (TFT-GAT) prediction model with Deep Double-Q Network (D3QN) scheduling optimization is proposed, validating the proposed framework’s robustness in spatiotemporal coupling, uncertainty representation, and real-time scheduling.
