A Novel Sine-Cosine Optimized LSTM for Smart Grid Load Forecasting

A smart grid derives its optimal value when it is paired with accurate load forecasting. In this paper, we present a novel hybrid approach that integrates the Sine-Cosine Optimization Algorithm (SCO) with Long Short-Term Memory (LSTM) networks in order to improve forecasting accuracy within smart grid environments. Though LSTM networks are proficient at capturing temporally dependent patterns as sequences, they suffer from one-dimensional hyperparameter tuning inefficiency concerning scope, which results in poor performance. This limitation can be overcome through SCO optimization of the learning rate, hidden units, and batch size in LSTM which increases the model's generalization and convergence rate. We evaluated the performance of the Sine-Cosine Optimized LSTM (SCO-LSTM) against real-life smart grid load datasets with seasonal and operational changes. Along with our proposed models, SCO-LSTM outperformed traditional LSTM, as well as all benchmark models, by a substantial margin in MAE, RMSE, and MAPE. The hybrid model also demonstrated outstanding robustness to drastic load changes, thus showcasing its potential for intelligent energy systems. This example shows the effectiveness and potential of bio-inspired optimization algorithms in deep learning for contemporary sophisticated power systems.