High-Resolution Solar Irradiance Prediction Using Variational Mode Decomposition and Entropy-Guided Multi-Expert Deep Learning Under Different Sky Conditions (Published)
Short-term solar irradiance prediction is investigated using regularly sampled measurements at one-minute or coarser resolutions, while decomposition-based learning on event-recorded, high-resolution measurements remains less explored. This study evaluates an entropy-guided multi-expert deep-learning framework combining variational mode decomposition (VMD), Sample Entropy, K-means clustering, and Neural Networks for high-resolution global horizontal irradiance (GHI) prediction. The framework uses the Natural Resources Canada high-resolution solar radiation dataset, comprising event-recorded measurements from two sensor networks in Quebec and Ontario. Eight observation days representing clear, overcast, variable, and very-variable conditions yield 164 sensor-day GHI series. Each series is decomposed into five modes using VMD; Sample Entropy is calculated for each mode, and K-means forms three complexity levels. Modes are assigned to multilayer perceptron (MLP), long short-term memory (LSTM), or one-dimensional convolutional neural network (1D-CNN) experts. Five mode-specific models are trained for each series under a fixed configuration, producing 820 trained model instances. Across the eight combinations, mean RMSE ranges from 0.57 to 6.40 W/m². Mean irradiance and relative error show a rank correlation of −0.933. The lowest-frequency mode (IMF1) was assigned to the MLP in all 164 series, while across all 820 IMF instances, 45.2% were assigned to the MLP, 43.0% to the LSTM, and 11.7% to the 1D-CNN. Because scaling and VMD precede chronological splitting and no baseline is included, results represent an empirical evaluation of the offline pipeline rather than strictly causal performance.
Keywords: 1D-CNN, High-resolution measurements, K-means clustering, LSTM, Sample entropy, Sensor Networks, Solar irradiance prediction, Variational mode decomposition, deep learning