A hyperband optimized Bi-LSTM model for electric demand forecasting in diversified power systems

Accurate prediction of electricity demand is vital for energy management and efficient operation of power systems. However, the dynamic variation of electric demand due to the rapid growth of unforeseen EV loads, battery charging, etc., poses challenges using the traditional methods of prediction. This paper presents an intelligent prediction model using a bi-directional long short-term memory (Bi-LSTM) network optimized with a hyperband strategy. The two-way design of the improved Bi-LSTM model understands the time-related patterns and makes its predictions more accurate by using hyperband techniques. Two sets of data are used to predict the electric demand using the optimized Bi-LSTM model, and its performance is compared with benchmark models like Extreme Gradient Boost (XGBoost) and deep neural networks (Standard LSTM and unoptimized Bi-LSTM). The optimized Bi-LSTM approach offers better accuracy in predicting electric demand in terms of reduced error metrics like mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), and correlation factor (R²). The performance in terms of error metrics during the training and testing phase indicates an immense drop in the case of the proposed optimized Bi-LSTM compared to the benchmark models. The MAPE for the optimized Bi-LSTM model in the current work is approximately 35 to 60 % lower compared to the benchmark models. However, the R² values have increased by 3 to 8 percentage points from 0.83 to 0.93 in benchmark models to around 0.91 to 0.99 in the proposed model. On the basis of the obtained performance, the proposed model is very trustworthy compared to other predictive models, as it provides higher accuracy of predictions.