An Optimised Deep Learning Model for Load Forecasting in Electric Vehicle Charging Stations

Accurate short-term load forecasting in the Electric Vehicle Charging Stations Network enhances power grid management. Existing methods often overfit the highly fluctuating energy consumption data from charging stations, creating a gap in developing accurate models. This paper tackles this challenge by proposing a ConvLSTM-BiLSTM, based encoder-decoder network, where convolutional layers are used to capture spatial trends along with recurrent layers for temporal dependencies. Furthermore, the model's hyperparameters are tuned using Levy Flight Particle Swarm Optimisation, enhancing its performance. The proposed model is evaluated on a publicly available Electric Vehicle Charging Stations dataset from Palo Alto City. The accuracy of the ConvLSTM-BiLSTM architecture with LFPSO optimisation surpasses that of conventional LSTM, BiLSTM models, and other encoder-decoder configurations. Significant improvements in RMSE, MAPE, and MSE were achieved, with reductions of around 37.14%, 62.13%, and 61.17%, respectively. The enhanced overall forecasting accuracy aids in better resource allocation and improves grid stability.