A deep neural network with two-step decomposition technique for predicting ultra-short-term solar power and electrical load

Solar penetration and energy consumption are rapidly increasing and transforming the modern power system due to population growth, urbanization, industrialization, and the electrification of transportation and heating. These changes bring uncertainties such as load demand dynamics and solar intermittent, which are being addressed through forecasting techniques to ensure a safe and reliable power system. However, the forecasting approach requires quality data by applying signal processing techniques to ensure accurate prediction. In this study, a two-step decomposition technique involving the denoising of the first intrinsic mode function (IMF) obtained from complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), integrated with a hybrid convolutional neural network-bidirectional long short-term memory (CNN-Bi-LSTM) model to enhance prediction accuracy has been proposed. The original time series was first decomposed into thirteen (13) IMFs, or sub-series by CEEMDAN. Frequency analysis was carried out on the generated IMFs to understand their individual characteristics and the frequency component of each IMF generated at the first stage of decomposition. Subsequently, high frequency component in the first IMF was removed differently using CEEMDAN or variational mode decomposition (VMD). The first and second stage decomposed signals are then input into the CNN-Bi-LSTM architecture, where the CNN efficiently captures local features and short-term dependencies, while the Bi-LSTM component excels at modeling long-term dependencies and temporal dynamics. Here, modeling results showed that two-step decomposition strategies improve the accuracy of the forecast significantly by reducing the uncertainties associated with solar power generation and electrical load demand. Our study showed that CEEMDAN-CEEMDAN based model is the best fit for datasets with low sampling frequency (electrical load dataset) while the CEEMDAN-VMD based model works well for datasets with both low (electrical load) and high (solar power) sampling frequencies. •Developed a two-step algorithm to enhance accuracy in time-series forecasting.•Proposed models show superior performance in forecasting solar power and load demands.•Explored the connections between signal frequencies and model prediction performance.