Integrating Ordinary Differential Equations With Sparse Attention for Power Load Forecasting

Accurate load forecasting plays an essential role in the measurement, monitoring, and control frameworks of modern power systems, particularly given the continuous influx of high-resolution data from advanced metering devices. Traditional forecasting methods often struggle due to the inherent nonstationarity and multiscale dynamics observed in these data streams. To address these challenges, this article introduces EvolvInformer, a novel long-sequence forecasting framework that integrates ordinary differential equations (ODEs) solver within a ProbSparse self-attention decoder architecture. The ODE module provides a physics-inspired, continuous-time representation of hidden state dynamics, enabling the model to capture subtle fluctuations and abrupt regime shifts commonly found in instrumented load profiles. Comprehensive experiments conducted on five large-scale power load datasets demonstrate that EvolvInformer achieves a 29.7% reduction in mean-squared error (mse) compared to state-of-the-art baseline models while preserving the logarithmic memory complexity characteristic of ProbSparse attention. Moreover, EvolvInformer consistently models both global trends and localized transient phenomena under stringent computational constraints, making it particularly suitable for embedded and edge-based metering applications. By effectively coupling continuous-time modeling via ODE with an efficient sparse attention mechanism for long-sequence forecasting, EvolvInformer provides a robust and scalable solution for measurement-centric load prediction tasks, with broad potential applications in adaptive energy management, grid load forecasting, and metering data quality assessment.