COMPARATIVE ANALYSIS OF MACHINE LEARNING LOAD FORECASTING TECHNIQUES

Load forecasting plays a critical role in energy management, and power systems, enabling efficient resource allocation, improved grid stability, and effective energy planning and distribution. Without accurate very short term load forecasting, utility management companies face uncertain load patterns, unrealistic prices, and poor infrastructure planning. This study aims to compare and evaluate different machine learning models and to recommend a model that is best suited for very short term load forecasting for cities as well as the most important predictors. To achieve this aim, quantitative research methodology, Case study research strategy were employed, and the research method used was quantitative research method. Data for this study was gotten from UCI data repository and several machine learning algorithms, including Artificial Neural Network-Long Short-Term Memory (ANN-LSTM), Support Vector Regression (SVR), Random Forests (RF), and Extreme Gradient Boosting (XGBOOST) were used to analyse the data. The dataset consists of 10 minutes interval of load consumption from January 2017 to December 2017 for 3 regions for the Moroccan city of Tatouan, weather variables, and other relevant factors that impact very short-term load forecasting. The performance of each machine learning technique in terms of its accuracy, feature importance, and model optimization is analysed. Model evaluation was done using various evaluation metrics, such as Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE). Model optimization was done using algorithms such as Grid Search and Hyperband. The results show the LSTM model as the most performant model in forecasting very shortterm load for cities with an MAE of 0.100, RMSE of 0.010, and MAPE of 4.25% . From a consensus of the various machine learning models, the most important predictors for forecasting very short-term load for cities are hour of the day, day of the year, general diffuse flows and temperature. Finally, model optimization was achieved using BayesSearchCV, GridSearchCV and Hyperband cross validation algorithms, sklearn timeseriessplit function with 10 folds cross validation to demonstrate how the performances of machine learning models could be improved using hyperparameter tuning and cross validation techniques