Deep Learning Framework for Joint Prediction of Energy Resources in an Industrial Building

Renewable energy sources like solar photovoltaic (PV) systems are inherently volatile, necessitating accurate short-term forecasting for efficient grid management, energy storage, and consumption optimization. This study proposes a deep learning framework using Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) models to simultaneously forecast PV generation (from roof and facade panels), on-site electrical consumption, and grid imports in an industrial building, relying solely on raw historical time-series data without meteorological or temporal features. The dataset, sourced from the Open Power System Data repository, comprises minute-by-minute measurements from a facility in Konstanz, Germany, preprocessed with Min-Max scaling and linear interpolation for gaps. The models were trained on ۸۰% of the data using a ۲۴-hour sequence length, with architectures featuring two recurrent layers (۵۰ units each), dropout (۰.۲), ReLU activation, Adam optimizer, and MSE loss. Early stopping and learning rate reduction callbacks prevented overfitting. Evaluation via MSE, RMSE, MAE, and R² on test data showed strong performance, with GRU achieving an average R² of ۰.۹۰۶, MAE of ۰.۰۳۴, MSE of ۰.۰۰۳۸, and RMSE of ۰.۰۶۲. Five-fold cross-validation confirmed model stability (mean R² ≈ ۰.۸۹ for both architectures), with GRU slightly outperforming LSTM. Results demonstrate the framework's ability to capture endogenous patterns, reducing data collection costs and enhancing applicability in legacy systems. Limitations include site-specific data, suggesting future enhancements via data augmentation, transfer learning, or hybrid models for broader generalizability. This approach supports sustainable energy management by minimizing fossil fuel reliance and operational costs.

کلیدواژه ها:

Photovoltaic Generation ، Neural Networks ، LSTM ، Short-Term Energy Forecasting ، Time-Series Analysis