Short-Term Load Prediction for Medium-Voltage Electricity Networks using Machine Learning: A Comparative Study

The rising uncertainties in electric load behaviour owing to human, technological and so-cio-economic events present a need to improve the accuracy and efficiency of current short-term load prediction (STLP) models. This paper compares the performance of four hybrid models for short-term Amp load prediction: Adaptive Neuro-Fuzzy Inference Sys-tem (ANFIS) integrated with Genetic Algorithm (GA) and Particle Swarm Optimisation (PSO), and Convolutional Neural Network (CNN) integrated with Long Short-Term Memory (LSTM) network and Extreme Gradient Boosting (XGB) machine. The models were trained and tested using historical data comprising hourly electrical Amp load ob-tained from a power utility substation in Kenya, and the corresponding weather data (temperature, wind speed, humidity) from January 2023 to June 2024. From the model testing results, both ANFIS-PSO and ANFIS-GA hybrid models show superior predictive accuracies with MAPE values of 4.519 and 4.636; RMSE of 0.3901 and 0.4024, and R2 scores of 0.9425 and 0.9391 respectively compared to CNN-LSTM and CNN-XGB models. The prediction across all models improved when the load data was pre-processed using Variational Mode Decomposition (VMD) technique. Nonetheless the hybrid ANFIS mod-els exhibited superior prediction accuracy, which is owed to their inherent adaptability to irregular data that enables them to capture the complex temporal patterns and non-linearities of Amp load well, thus making them more suitable for short-term load prediction problems.