Load Forecasting Research Articles

A Shapley value-based dynamic ensemble framework for short-term load forecasting of industrial consumers

An adaptive multi-seasonal ARIMA approach for domestic hot water load forecasting: A pilot study

A hybrid deep learning model for load forecasting of electric vehicle charging stations using time series decomposition

Accurate load forecasting of central air conditioning (CAC) systems is crucial for enhancing energy efficiency and minimizing operational costs. However, the complex nonlinear correlations among meteorological factors, water system dynamics, and cooling demand make this task challenging. To address these issues, this study proposes a novel hybrid forecasting model termed IWOA-BiTCN-BiGRU-SA, which integrates the Improved Whale Optimization Algorithm (IWOA), Bidirectional Temporal Convolutional Networks (BiTCN), Bidirectional Gated Recurrent Units (BiGRU), and a Self-attention mechanism (SA). BiTCN is adopted to extract temporal dependencies and multi-scale features, BiGRU captures long-term bidirectional correlations, and the self-attention mechanism enhances feature weighting adaptively. Furthermore, IWOA is employed to optimize the hyperparameters of BiTCN and BiGRU, improving training stability and generalization. Experimental results based on real CAC operational data demonstrate that the proposed model outperforms traditional methods such as LSTM, GRU, and TCN, as well as hybrid deep learning benchmark models. Compared to all comparison models, the root mean square error (RMSE) decreased by 13.72% to 56.66%. This research highlights the application potential of the IWSO-BiTCN-BiGRU-Attention framework in practical load forecasting and intelligent energy management for large-scale CAC systems.

Conventional power load forecasting frameworks face limitations in dynamic spatial topology capture and long-term dependency modeling. To address these issues, this study proposes a hybrid GAT-CNN-LSTM architecture for enhanced short-term power load forecasting. The model integrates three core components synergistically: Graph Attention Network (GAT) dynamically captures spatial correlations via adaptive node weighting, resolving static topology constraints; a CNN-LSTM module extracts multi-scale temporal features—convolutional kernels decompose load fluctuations, while bidirectional LSTM layers model long-term trends; and a gated fusion mechanism adaptively weights and fuses spatio-temporal features, suppressing noise and enhancing sensitivity to critical load periods. Experimental validations on multi-city datasets show significant improvements: the model outperforms baseline models by a notable margin in error reduction, exhibits stronger robustness under extreme weather, and maintains superior stability in multi-step forecasting. This study concludes that the hybrid model balances spatial topological analysis and temporal trend modeling, providing higher accuracy and adaptability for STLF in complex power grid environments.

Abstract This study develops a unified framework for accurate load forecasting and planning optimization in low-voltage distribution networks, addressing challenges from load volatility and renewable integration. We propose an Adaptive DeepTemporalNet model, which incorporates a Dynamic Feature Capture Module (DFCM) and an Adaptive Weight Adjustment Module (AWAM), enabling simultaneous modeling of short- and long-term load dependencies while dynamically reweighting critical features. Experiments on a three-year dataset show that Adaptive DeepTemporalNet reduces mean squared error (MSE) by 32% compared to LSTM and over 50% compared to ARIMA, achieving an average mean absolute percentage error (MAPE) of 3.5%. Furthermore, by integrating load forecasting with a stochastic chance-constrained optimization framework, the proposed approach lowers total planning costs by about 15% and improves voltage compliance rates by up to 12 percentage points. These results demonstrate both theoretical contributions—advancing adaptive temporal modeling and stochastic optimization—and practical benefits for grid operators seeking stable, cost-effective, and renewable-friendly network operation.

Virtual Power Plant (VPP) is capable of aggregating and intelligently coordinating diverse distributed energy resources, among which the accuracy of load forecasting is a key factor in ensuring their regulation capability. To address the periodicity and complex nonlinear fluctuations of electricity load data, this study introduces a Cyclic Order Mapping (COM) encoding method, which maps weekly and intraday sequences into continuous ordered variables on the unit circle, thereby effectively preserving load periodic features. On the basis of the COM encoding, a novel forecasting model is proposed by integrating Bidirectional Long Short-Term Memory (BiLSTM) networks, an efficient self-attention mechanism, and the Kolmogorov–Arnold Network (KAN). This model is termed BiLSTM-Att-KAN. Comparative and ablation experiments were conducted to assess the scientific validity and predictive accuracy of the proposed approach. The results confirm its superiority, achieving a Root Mean Square Error (RMSE) of 141.403, a Mean Absolute Error (MAE) of 106.687, and a coefficient of determination (R2) of 0.962. These findings demonstrate the effectiveness of the proposed model in enhancing load forecasting performance for VPP applications.

Electricity load forecasting is crucial for people's daily lives. It is critical to identify an appropriate model for load prediction, as it is paramount to achieving reliable results. This article aims to explore methods for predicting short-term electricity load. The auto-regressive integrated moving average (ARIMA) model is applied to analyze the data, which consists of 2,182 consecutive hourly load values starting from 0:00 on March 1st, 2003. Four variables affecting electricity load are selected. Seasonal influences are also taken into account, and a seasonal ARIMA approach is adopted to mitigate bias caused by seasonality. To evaluate the effectiveness of the method, the Ljung-Box Q-test is performed on the residuals of the forecasted values. The results indicate that the Seasonal ARIMA model achieves the best fit, the short-term prediction is reliable. The ARIMA model only requires historical data to generate relatively accurate predictions, eliminating the need for extensive datasets, and the process is relatively simple. Overall, short-term electricity load can be explained by both historical load values and past errors.

Accurate short-term air conditioning load forecasting is very important for controlling air conditioning energy consumption, and is also a requisite for realizing intelligent control of air conditioning units. This paper proposes a short-term load forecasting framework for air conditioning based on the concept of model stacking, which combines six mature machine learning models, including Lasso regression, Ridge regression, Random Forest, Support Vector Regression, eXtreme Gradient Boosting and Long Short-Term Memory, and trains new prediction models through model stacking. In order to realize the short term forecast function of air conditioning load, this paper presents a complete forecast framework, which includes operational procedures for feature screening, algorithm hyperparameters optimization, and cross stacking of prediction models. A real air conditioning system is employed for prediction analysis, the prediction results showed that 28 out of 36 control simulations achieved better prediction accuracy, with an average increase in R2 of 6.4%. Notably, simpler submodels in the meta-model yield better results in model stacking, whereas complex coupling models as the meta-model may degrade performance. The proposed framework not only improves prediction accuracy but also maintains reasonable training time and resource requirements, making it practical for real-world applications. These findings provide insights into implementation of model stacking and selecting the meta-model for short-term air conditioning load forecasting.

In smart metering systems, data loss often occurs due to sensor failures, communication delays, and equipment maintenance, affecting the accuracy of power data analysis. This study proposes an box-meter integrated metering device that supports localized data imputation and combines it with deep learning models for further research. We compared the imputation performance of different deep learning models–including DLinear, TimesNet, and iTransformer–under varying missing rates. Experimental results show that TimesNet achieves optimal imputation performance across diverse missing scenarios. The device is capable of deploying deep learning models and integrates a raw analog signal acquisition interface, thereby reducing data loss at the source and enhancing data continuity and real-time availability. This approach improves data quality and timeliness, providing a solid data foundation for power system tasks such as intelligent scheduling and load forecasting.

Accurate forecasting of residential heating loads is crucial for guiding heating system control strategies and improving energy efficiency. In recent years, research on heating load forecasting has primarily focused on continuous district heating systems, and it often struggles to cope with the abrupt load fluctuations and irregular on/off schedules encountered in intermittent heating scenarios. To address these challenges, this study proposes a hybrid convolutional long short-term memory (ConvLSTM) model that replaces the conventional batch normalization layer with a Dynamic Tanh (DyT) activation function, enabling dynamic feature scaling and enhancing responsiveness to sudden load spikes. An improved channel–temporal attention mechanism, CBAM(T), is further incorporated to deeply capture the spatiotemporal relationships in multidimensional data and effectively handle the uncertainty of heating start–stop events. Using data from two heating seasons for households in a residential community in Dalian, China, we validate the performance of ConvLSTM-DyT-CBAM(T). The results show that the proposed model achieves the best predictive accuracy and strong generalization, confirming its effectiveness for intermittent heating load forecasting and highlighting its significance for guiding demand-responsive heating control strategies and for energy saving and emissions reduction.

During the peak load period, there is a high level of imbalance between power supply and demand, which has become a critical challenge, leading to higher operational costs for power grids. To improve the accuracy of peak load forecasting, this study introduces a novel approach based on Extreme Gradient Boosting Trees (XGBoost) and Multiple Linear Regression (MLR) for daily peak load prediction. The proposed methodology first employs an improved version of the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN) algorithm to decompose the raw load data, subsequently reconstructing each Intrinsic Mode Function (IMF) into high-frequency and stationary components. For the high-frequency components, XGBoost serves as the base predictor within a Bagging-based ensemble structure, while the Sparrow Search Algorithm (SSA) is employed to optimize hyperparameters automatically, ensuring efficient learning and accurate representation of complex peak load fluctuations. Meanwhile, the stationary components are modeled using MLR to provide fast and reliable estimations. The proposed framework was evaluated using actual daily peak load data from Western Inner Mongolia, China. The results indicate that the proposed method successfully captures the peak characteristics of the power grid, delivering both robust and precise predictions. When compared to the baseline model, the RMSE and MAPE are reduced by 54.4% and 87.3%, respectively, underscoring its significant potential for practical applications in power system operation and planning.

Energy Management (EM) in hybrid Microgrids (MGs) is essential for coordinating Renewable Energy Sources (RESs) and Hybrid Energy Storage Systems (HESSs) to ensure Power Quality (PQ), stable operation, and efficient power flow. Existing optimization–prediction approaches often address these issues in isolation or require high computational overhead, limiting their real-time applicability. To overcome these challenges, this paper proposes a novel dual-optimization framework combining the Artificial Lemming Algorithm (ALA) with the Temporal Kolmogorov–Arnold Network (TKAN), referred to as ALA-TKAN. Unlike conventional methods, ALA-TKAN integrates metaheuristic-based optimization of power flow and HESS scheduling with sequence-aware forecasting of load and renewable generation, enabling proactive and coordinated EM under dynamic conditions. Implemented in MATLAB, the proposed method demonstrates superior performance compared with state-of-the-art techniques such as PDO-MACNN, BWO, PSO, ANN, and MRA-FLC, achieving minimal power loss (2.9 MW), highest efficiency (99.2%), lowest energy cost (0.8 $/Wh), and reduced THD (1.4%). These results confirm the novelty and practical potential of ALA-TKAN as a unified, computationally efficient strategy for PQ enhancement and reliable operation of hybrid MGs.

As an important research direction for the future innovation and development of the energy industry, integrated energy system (IES) requires short-term load forecasting as the decision-making basis. A short-term load forecasting model based on improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) maximum information coefficient (MIC) transformer is proposed to address the problem of high volatility and strong randomness in IES short-term load. The model effectively reduces the complexity of time series and alleviates the impact of modal aliasing by ICEEMDAN. Perform correlation analysis and reconstruction of the decomposed intrinsic mode function using MIC to improve the coupling relationship between multivariate loads. Finally, Transformer model based on self-attention mechanism is used to predict each component and obtain the final prediction result. The horizontal comparison experiment and vertical ablation experiment were conducted on the model using the IES dataset from Arizona State University in the United States. The results showed that the predictive performance of the model was improved to a certain extent and it had a certain degree of application prospects.

This study proposes a novel Time series Decomposition and Attention Graph Neural Network (TDAGNN) that integrates temporal decomposition and attention mechanisms to address the challenges of complex spatiotemporal coupling and abrupt phase transitions in load forecasting for high-rise building construction. A Dual Time series Decomposition Convolutional Neural Network (DTDCNN) is employed to extract temporally dependent features from intricate high-altitude construction load data. To effectively capture the heterogeneous and dynamic characteristics of load fluctuations, a Multi-head Interactive Attention (MIA) module is introduced, enabling interactive learning between original and locally enhanced features. Furthermore, a Self-scaling Dynamic Diffusion Graph Neural Network (SDDGNN) is incorporated to model spatial dependencies while mitigating scale distortion commonly encountered in graph-based methods. Experimental evaluations on the expanded BIM-SHMC dataset demonstrate that the proposed framework achieves state-of-the-art performance with peak load error rate (PPER) of 8.72 ± 0.23%, power fluctuation matching degree (PFMD) of 92.17 ± 0.41%, resource scheduling alignment (RSA) of 94.32 ± 0.38%, and phase transition detection accuracy (PHA) of 91.25 ± 0.35%, representing an average improvement of 13.3% over the next-best model (STGCN).

Accurate electricity load forecasting is crucial for power system planning, reliability, and sustainability, enabling more efficient markets and reduced greenhouse gas emissions. This study leverages deep learning algorithms, specifically bidirectional recurrent neural networks, to develop a unified model for predicting one day-ahead electricity demand for the entire year of 2023. The model's performance was evaluated on a monthly basis, allowing for a detailed assessment of its forecasting capabilities across different time periods. Four neural network algorithms were compared: Long Short-Term Memory (LSTM), Bidirectional LSTM, Gated Recurrent Unit (GRU), and Bidirectional GRU. The GRU model demonstrated superior performance, achieving an R-squared value of 0.8526 in October and a Mean Absolute Percentage Error (MAPE) of 2.34% in March. These results highlight the potential of the proposed model as an effective tool for electricity demand forecasting, supporting the integration of renewable energy sources and enhancing grid resilience.

IntroductionWith increasing uncertainties on both the generation and load sides in power systems, ultra-short-term load forecasting (USTLF) and risk assessment have become crucial for ensuring the secure and optimal operations of power systems, especially in distribution networks.MethodsThis paper proposed a probabilistic load forecasting method that integrates variational mode decomposition (VMD) with an improved deep autoregressive probabilistic forecasting (DeepAR) model. VMD reduces the non-stationarity of the load sequence, and a future feature enhancement mechanism was introduced to improve the accuracy under multi-step predictions. Based on the proposed method, an integrated assessment framework covering voltage deviations and transformer overload risks was constructed. Exponential aggregation functions and nonlinear normalization methods were utilized to evaluate the combined risk index with multidimensional risk indicators with different units.ResultsCase studies demonstrated that the proposed VMD with the improved DeepAR model improved the accuracy of load forecasting over traditional models.DiscussionMoreover, the proposed risk assessment method can provide quantitative and systematic early risk-warning support for distribution network operations and decision-making.

Load forecasting is an asset for sustainable building energy management, as accurate predictions enable efficient energy consumption and con- tribute to decarbonisation efforts. However, data-driven models are often limited by dataset length and quality. This study investigates the effectiveness of transfer learning (TL) for load forecasting in office buildings, with the aim of addressing data scarcity issues and improving forecasting accuracy. The case study consists in a group of eight virtual buildings (VB) located in Porto, Portugal. VB A2 serves as pre-trained base model to transfer knowledge to the remaining VBs, which are analysed in varying degrees of data availability. Our findings indicate that TL can significantly reduce training time, for up to 87%, while maintaining accuracy levels comparable to those of models trained with full dataset, and exhibiting superior performance when com- pared to models trained with scarce data, with average RMSE reduction of 42.76%.

In the context of climate change and energy transition, the growing frequency of extreme weather events threatens the safety and stability of power systems. Given the limitations of existing research on load characteristic analysis and load forecasting during extreme weather events, this paper proposes a load-integrated forecasting model that accounts for extreme weather. First, an improved power load clustering method is proposed, combining Kernel PCA for nonlinear dimensionality reduction and an enhanced k-means algorithm, enabling both qualitative analysis and quantitative representation of load characteristics under extreme weather. Second, an optimal combination forecasting model is developed, integrating improved SVM and enhanced LSTM networks. Building upon the improved power load clustering algorithm, a load-integrated forecasting model considering extreme weather is established. Finally, based on the proposed load-integrated forecasting model, a time-series production simulation model considering extreme weather is constructed to quantitatively analyze the power and electricity balance risks of the system. Case studies demonstrate that the proposed integrated forecasting model can effectively analyze load characteristics under extreme weather and achieve more accurate load forecasting, which can provide guidance for the planning and operation of new power systems under extreme weather conditions.

Building energy flexibility is essential for integrating renewables, optimizing energy use, and ensuring grid stability. While renewable and storage systems are increasingly used in buildings, poorly designed storage strategies often cause supply-demand mismatches, and a comprehensive, indicator-based assessment approach for quantifying flexibility remains lacking. Therefore, this study designs customized energy storage regulation strategies and constructs a comprehensive energy flexibility assessment scheme to address key issues in supply-demand coordination and energy flexibility evaluation. LSTM and Rolling-XGB methods are used to predict building energy consumption and PV generation, respectively. Based on battery safety constraints, a data-driven battery energy storage system (BESS) model simulates battery behavior to evaluate and compare building energy flexibility under two scenarios: (1) uncoordinated PV-BESS, and (2) coordinated PV-BESS with load forecasting. A practical validation was conducted using a net-zero-carbon building as the case study. Simulation results show that the data-driven BESS model improves building energy flexibility and reduces electricity costs through optimized battery sizing, tailored storage strategies, and consideration of local time-of-use tariffs. In the case study, local energy coverage reached 62.75%, surplus time increased to 34.77%, and costs were cut by nearly 40% compared to the PV-only scenario, demonstrating the significant benefits brought by the proposed BESS model that integrates load forecasting and PV generation prediction features.