TLDR: A new deep learning method, CEEDMA-Multi_nets-EQN, significantly improves ultra-short-term solar power forecasting. It achieves this by decomposing solar data into high and low-frequency components, using specialized neural networks (CNN, iTransformer, BiLSTM) for each, fusing features with multi-head attention, and employing an Evidential Quantile Network (EQN) with a novel width penalty for accurate probabilistic predictions. Experiments show superior accuracy, stability, and resource efficiency compared to existing models.
Solar power is a cornerstone of the global shift towards green energy, but its inherent variability poses significant challenges to grid stability and energy management. Accurately predicting solar power output, especially for very short periods (ultra-short-term forecasting), is crucial for integrating this renewable energy source effectively. This challenge is precisely what a recent research paper, “Ultra-short-term solar power forecasting by deep learning and data reconstruction,” addresses.
Authored by Jinbao Wang, Jun Liu, Shiliang Zhang, and Xuehui Ma, this paper introduces a novel deep learning approach that combines advanced data decomposition and reconstruction techniques with sophisticated neural networks to achieve highly accurate and reliable solar power predictions. The core idea is to better understand and utilize the complex patterns within solar power data, which are influenced by various factors like cloud movements and daily solar cycles.
Traditional forecasting methods often struggle with the multi-scale temporal features present in solar power data, leading to lower accuracy or computational inefficiencies. The proposed method, named CEEDMA-Multi_nets-EQN, tackles these limitations by first breaking down the raw solar power data into more manageable components. It uses a technique called Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) to decompose the data into several Intrinsic Mode Functions (IMFs) and a residual component. This process effectively reduces the non-stationary nature of the original signal, making it easier to analyze.
A key innovation of this research lies in its data reconstruction strategy. Instead of directly using the numerous IMFs, the authors reconstruct them into just two main components: a high-frequency component and a low-frequency component. The high-frequency component captures rapid changes, such as sudden cloud cover or dew evaporation, while the low-frequency component reflects slower, more predictable patterns like diurnal cycles and seasonal variations. This reconstruction significantly improves the model’s ability to generalize across different datasets, as the number of IMFs can vary widely depending on data complexity.
Once the data is reconstructed, the method employs a specialized multi-network architecture for feature extraction. A Convolutional Neural Network (CNN) is used for the high-frequency component, adept at capturing local features and rapid changes. An iTransformer, known for handling long-term dependencies, processes the low-frequency component. Additionally, meteorological data (like irradiance, temperature, humidity, and visibility, identified through correlation analysis) are fed into a Bidirectional Long Short-Term Memory (BiLSTM) network to extract relevant weather-related features. These diverse features are then intelligently combined using a multi-head attention mechanism, allowing the model to focus on the most important aspects for prediction.
For the final prediction, the fused features are passed to an Evidential Quantile Network (EQN). This network is designed not only to provide precise point predictions but also to quantify the uncertainty associated with those predictions, offering a probabilistic forecast. A crucial enhancement to the EQN is the inclusion of a “width penalty term” in its loss function. This penalty discourages overly wide prediction intervals, ensuring that the uncertainty ranges provided are both accurate and practically useful, preventing the model from becoming “overconfident” or generating meaningless broad intervals.
The effectiveness of the CEEDMA-Multi_nets-EQN method was rigorously tested using data from a rooftop solar project at the Hong Kong University of Science and Technology, involving four different types of PV equipment. Extensive experiments, including ablation studies, demonstrated that both the data reconstruction and the width penalty term significantly improve performance. When compared against five baseline models, the proposed method consistently showed superior prediction accuracy, stability, and generalization across all datasets. For instance, it achieved optimal results for metrics like R2, CRPS, and WS, indicating a high degree of fit with true values and more reliable probabilistic forecasts.
Furthermore, the research also evaluated the computational resource consumption of the model. The proposed method exhibited efficient CPU and GPU utilization, suggesting that it can be deployed on devices with lower computational power while still delivering excellent prediction performance. This makes it a practical solution for real-world energy management systems.
Also Read:
- Enhancing EV Charging Predictions: A Hybrid Deep Learning Approach
- ScenGAN: Enhancing Renewable Energy Forecasting with Advanced Uncertainty Modeling
In conclusion, the CEEDMA-Multi_nets-EQN method represents a significant advancement in ultra-short-term solar power forecasting. By intelligently decomposing and reconstructing data, employing specialized deep learning architectures for different data components, and incorporating a novel uncertainty quantification mechanism, it offers a robust, accurate, and stable solution for managing the intermittency of solar energy. For more detailed information, you can read the full research paper here.
