Sep 2026· IEEE Transactions on Smart Grid· Vol 17, pp. 4008-4020· 0 citations· 52 references
Abstract
The growing integration of renewable energy sources (RESs), such as photovoltaic (PV) and wind (WD), has significantly increased the variability of net load (NL), posing critical challenges on the net load forecasting. In this paper, a novel physics-informed series-aware graph Transformer (PISAGT) model is proposed for the net load forecasting, which synergistically combines the advantages of both direct and indirect forecasting methods to achieve superior forecasting performance. Firstly, a physics-informed loss module (PILM) is proposed, which introduces the physical characteristics of the net load into the loss function to incorporate physical knowledge in the gradient descent optimization, thereby enhancing the model’s generalization capability for more stable and trustworthy predictions. Secondly, a variable-scale patch embedding method segments time series into subsequence-level patches and transforms them into 2D representations, which enables the forecast model to simultaneously capture local temporal patterns across diverse forecasting horizons. Thirdly, a series-aware graph learning mechanism (SAGLM) which includes graph token encoding, adaptive adjacency matrix learning, and graph convolution, is proposed to integrate graph neural networks (GNNs) with Transformer within a synergistic framework to comprehensively capture intra-series and inter-series dependencies in net load datasets. Finally, case studies on the real-world datasets demonstrate that the PISAGT model can achieve a 14.4% improvement in prediction accuracy compared with direct forecasting methods, showing its effectiveness in net load forecasting.
Tourist flow modeling and prediction in scenic areas is highly complex, exhibiting significant spatiotemporal dependence and being influenced by various external factors. Traditional models struggle to simultaneously characterize complex spatiotemporal relationships and integrate diverse external information, posing challenges for intelligent resource management and dynamic information scheduling in large-scale networked systems. To address this, this paper proposes 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. The framework integrates heterogeneous data such as historical traffic, weather, and social media, achieving adaptive time-varying spatial embedding through a graph attention network (GAT) while capturing long- and short-term dependencies using a Temporal Fusion Transformer (TFT) for both point and quantile prediction. The PSO algorithm performs global optimization of the TFT-GAT hyperparameters, and the resulting prediction outputs together with uncertainty estimates are incorporated into the D3QN to realize closed-loop online capacity allocation based on reinforcement learning. Such a data-driven spatiotemporal modeling strategy also provides methodological insights for dynamic information fusion and adaptive resource scheduling in intelligent electromagnetic sensing and communication environments. Experiments conducted at the Forbidden City in Beijing demonstrate high prediction accuracy with a mean absolute error of 1.5–2.6 people/hour, an average quantile coverage exceeding 82%, and a response time of 140.6 ms. The incorporation of exogenous factors, particularly holidays, improves prediction performance by 18.4%, validating the proposed framework’s robustness in spatiotemporal coupling, uncertainty representation, and real-time scheduling.
X. N. Xie, W. L. Zhang· Advanced Electromagnetics· 0 citations
The high penetration of renewable energy and the large-scale integration of electric vehicles have significantly enhanced the nonlinearity and non-stationarity of the smart grid load sequence. Traditional time series models struggle to meet the prediction requirements in complex scenarios. This paper focuses on three kinds of deep learning load forecasting methods: recurrent neural network variants, CNN hybrid architectures, and Transformers, and systematically conducts a comparative analysis of them from three dimensions: prediction accuracy, long-term dependency modeling ability and computational efficiency. The results show that the three types of architectures each have distinct advantages, and there is no universal optimal solution: the recurrent variant balances accuracy and efficiency in stationary short-term prediction, the CNN hybrid architecture is more robust in strong coupling scenarios such as extreme weather, and the Transformers' trend modeling ability is outstanding but the computational cost for long sequences is large. The pros and cons of each method are highly dependent on the stationarity of the specific scenario, the forecasting horizon, and resource constraints. The model selection must be based on the task characteristics rather than the pursuit of general solutions.
Qiuhui Jiang· Applied and Computational En...· 0 citations
Accurate power load forecasting is critical for the efficient operation of industrial microgrids. However, raw meteorological and consumption data typically exhibit non-stationary characteristics, complicating the hyperparameter tuning of deep learning models, and subsequently degrading the prediction accuracy of these frameworks. To address the aforementioned challenges, a new hierarchical forecasting structure denoted as INRBO-SSA-LSTM is proposed in this paper. First, Pearson correlation analysis is employed for feature reduction, identifying the four main factors to mitigate the dimensionality curse. Building upon this foundation, a refined Newton-Raphson-Based Optimizer (INRBO) is introduced, integrating a cosine adaptive t-distribution perturbation, a boundary-aware non-uniform steering scheme, and a fitness-aware hybrid perturbation mechanism. Evaluated against the CEC2022 benchmark suite, comprehensive evaluations reveal that the INRBO demonstrates superior global exploration and local refinement capabilities compared to baseline algorithms when assessed on the CEC2022 benchmark suite for foundational optimization performance. Furthermore, rigorous testing on the CEC2017 suite across 10, 30, and 50 dimensions successfully validates its exceptional robustness and search capabilities in high-dimensional spaces. INRBO functions as a dual-stage optimizer within the proposed framework; in the initial phase, it dynamically calibrates the parameters of Singular Spectrum Analysis (SSA) to extract deterministic load patterns, achieving a maximum signal-to-noise ratio of 15.87 dB; in the second phase, it optimizes the global hyperparameters of the Long Short-Term Memory (LSTM) network. Validated using actual industrial microgrid data in Jiangsu Province, China, the proposed method significantly outperforms traditional baseline models across all indicators; specifically, the prediction error (RMSE = 10.9764, MAPE = 3.7866%) is substantially minimized, and the coefficient of determination (R2 = 0.9741) is highly optimal. This adaptable framework effectively accommodates temporal demand variations, offering a robust foundation for the advancement of intelligent power management technology.
Jinming Luo, Fu Chen, Lingshan Kong et al.· Electronics· 0 citations
Electricity load forecasting is essential for balancing energy supply and demand, reducing energy waste, and maintaining power grid stability. Accurate forecasts enable power utilities to optimize energy dispatch and mitigate the risk of supply shortages or outages. However, conventional forecasting methods often struggle to capture highly nonlinear local fluctuations in electricity consumption and long-term temporal dependencies. To address these challenges, this study proposes MSCNN-ResLSTM, a hybrid model for multi-step electricity load forecasting. The proposed model integrates Multi-Scale Convolutional Neural Networks (MSCNNs) to extract local time-series features at multiple temporal scales, residual blocks (ResBlocks) to enhance feature representation through residual connections, and Long Short-Term Memory (LSTM) networks to model long-range temporal dependencies. To comprehensively evaluate its effectiveness, a cross-paradigm experimental framework is established in which MSCNN-ResLSTM is compared with seven representative benchmark models from three methodological categories: traditional machine learning (Extreme Gradient Boosting-XGBoost), classical recurrent and convolutional neural networks (LSTM, Temporal Convolutional Network-TCN, CNN-LSTM, MSCNN-LSTM, and Direct LSTM (Seq2Seq)), and self-attention-based architectures (Transformer). Experimental results show that MSCNN-ResLSTM achieves higher forecasting accuracy and greater stability across the full 24-step prediction horizon, consistently outperforming all competing baselines while effectively suppressing recursive error propagation.
Reliable short-term prediction of electricity consumption is essential for dispatching power grids and incorporating renewable sources. This work presents an interpretable forecasting system that merges heterogeneous data streams with an attention-augmented PatchTST backbone. Real-world meteorological records from NOAA are combined with electricity demand data, and we design time-adaptive correlation descriptors alongside a binary extreme-weather flag. A one-dimensional Coordinate Attention (CA) mechanism, originally developed for image analysis, is integrated into the PatchTST encoder to enable channel-wise reweighting. Using the UCI Electricity Load dataset for customer MT_329 (2011-2012), our CA-PatchTST attains a MAE of 16.46, surpassing LSTM (18.40) and the unmodified PatchTST (17.59). Ablation tests indicate that incorporating meteorological information is the primary driver of accuracy gain, while the CA module contributes moderate improvement and, crucially, offers visualizable attention maps that clarify the model's behavior under severe weather. The fusion of real NOAA observations lifts forecasting performance by over 50% relative to load-only models. The dynamic correlation features effectively capture seasonal reversals in load–weather dependency across months. Overall, the proposed approach achieves competitive results (MAE 16.46) on the benchmark.
Bohan Zhang· Applied and Computational En...· 0 citations
Accurate photovoltaic (PV) power forecasting is essential for reliable microgrid operation, efficient energy dispatch, and improved utilization of renewable energy resources. Existing forecasting methods often have limited capacity to represent the nonlinear relationships between meteorological conditions and PV power output. They also tend to underrepresent the temporal dynamics of PV generation and the physical principles governing photovoltaic energy conversion. To address these limitations, this study proposes a hybrid forecasting framework, CNN-LSTM-PINNs, that integrates Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, and Physics-Informed Neural Networks (PINNs). In the proposed framework, CNNs extract spatial dependencies among multivariate meteorological variables, LSTM networks capture temporal dependencies in PV generation, and PINNs incorporate soft physical constraints derived from photovoltaic energy conversion mechanisms. The proposed model is evaluated using publicly available datasets from three large-scale PV power stations in China, with observations recorded at 15-min intervals. The empirical results show that CNN-LSTM-PINNs outperform the conventional CNN-LSTM benchmark across the primary station-level datasets. Relative to the benchmark model, the proposed framework reduces Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) and improves the coefficient of determination (R2). These results indicate that embedding physical constraints into data-driven forecasting models can improve PV power prediction accuracy. The model also shows stronger robustness and generalization performance under heterogeneous operating conditions, although its effectiveness is contingent on relatively stable data distributions. Feature-importance analysis further indicates that global horizontal irradiance (GHI) and irradiance-derived variables are the most informative predictors of PV power output. Overall, this study provides a physics-informed hybrid modeling approach for high-resolution PV power forecasting in microgrid applications.
Jiabo Gou, Xiaoqiao Liao, Sheng Li et al.· Energies· 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.
