A Hybrid CNN-LSTM Model with Signal Decomposition for Multi-Step Electrical Load Forecasting

Accurate power load forecasting is a cornerstone for the reliable operation and economic dispatch of modern power grids, particularly as the integration of Variable Renewable Energy continues to rise. However, the intrinsic non-linearity, randomness, and intricate multi-seasonal behavior of load data pose formidable challenges for traditional models, which often fail to simultaneously capture long-range dependencies and high-frequency variations. To address these challenges, this paper proposes a novel hybrid deep learning model that leverages a decomposition-driven, dual-stream architecture, synergizing time domain and frequency domain learning. The denoised full signal, enriched with calendar-based encodings, is modeled by a Long Short-Term Memory (LSTM) network to extract holistic temporal patterns. Concurrently, Seasonal-Trend decomposition using Loess (STL) isolates seasonal and residual components, which are further processed via Fast Fourier Transform (FFT) and analyzed by a one-dimensional Convolutional Neural Network (CNN), specialized in spectral pattern recognition. The feature vectors from two streams are fused and predicted through a Multi-Layer Perceptron (MLP) to generate multi-step forecasts. Experimental results on Belgian power load data (2020–2024) demonstrate the superior predictive accuracy and robustness. This research contributes a novel and interpretable modeling structure by strategically allocating forecasting subtasks—contextual modeling to LSTM, spectral detection to CNN—providing a robust solution for complex time series forecasting in energy systems.