Interpretable Hybrid Deep Learning for Real-Time Multi-Horizon Forecasting of Load, Solar, and Wind Generation Using Multi-Source Energy and Weather Data

Abstract

The proposed study introduces synthesizable hybrid deep learning architecture in real time multi-horizon load, solar, and wind power generation prediction through combination of multi-source of meteorological and power system observations. To meet the increasing demand of precise and explainable energy predictions, the given model integrates empirical wavelet transform (EWT) to break down the signal down into small pieces with advanced neural networks, including LSTM, CNN, and Transformers, and exalted with SHAP to be explainable. Public datasets were also based on the weather (temperature, humidity, wind speed) and energy generation/load that were used in various geographic areas. The comparative benchmarking between the traditional (Random Forest, XGBoost) and standalone deep learning models showed that the hybrid model performed better and was rated as 5-12 times worse in terms of RMSE and R-2 metrics. In particular, the accuracy of the predictions increased as well as the forecasting accuracy improved (0.82 in the traditional models compared to 0.91 in the hybrid with weather integration) and the inference time was minimized to be deployed in real-time. Transfer learning should be used to adapt to the domain so that fine-tuning can be done on target regions and that the increase in accuracy can be retained in the face of regional changes in data. Analysis of interpretability revealed that weather conditions including temperature and wind velocity contributed greatly to predictions, which increased the confidence of stakeholders. The findings show the possibilities of interpretable hybrid models in the smart grid management and point to the future directions like the deployment of edges, real-time updates based on IoT, and policy-level integration.