Short-term enterprise electricity consumption forecasting in the energy data space using ISSA-VWELM

Abstract
Accurate short-term enterprise electricity-consumption forecasting is important for energy management and operational decision-making in the energy data space. However, enterprise-level load series are often affected by nonlinear dynamics, heterogeneous production behavior, irregular fluctuations, and noise, making accurate and stable forecasting difficult. To address this issue, this study proposes ISSA–VWELM, a hybrid forecasting model that combines an improved sparrow search algorithm (ISSA) with a sample-level variance-weighted extreme learning machine (VWELM). VWELM estimates output weights through a weighted least-squares scheme, where adaptive sample weights are determined by local residual variances from a preliminary ELM model to reduce the influence of unstable or noisy observations. ISSA is used to optimize the input weights and hidden-layer biases of VWELM, while good point set initialization, the golden sine mechanism, and Levy flight are introduced to improve the search ability of SSA. The proposed method is evaluated through benchmark-function optimization, noisy nonlinear SinC regression, and real-world forecasting using daily electricity-consumption data from 38 enterprises in the Jiangsu Energy Data Space. Results show that ISSA improves the convergence of SSA, and ISSA–VWELM achieves stronger robustness than several ELM-based baselines under mixed-noise conditions. In the real-world forecasting task, ISSA–VWELM obtains the best performance, with an RMSE of 130.9942, a MAPE of 0.63%, and an MAE of 95.8300, reducing these errors by 12.65%, 8.70%, and 6.31% compared with VWELM. These findings demonstrate the effectiveness of ISSA–VWELM for heterogeneous and noisy enterprise electricity-consumption forecasting.