Short-term electricity load forecasting using classification and regression tree and deep learning models

Short-term load forecasting (STLF) is an important section for the planning and operation of electricity generators to arrange the units optimally. Nonetheless, accurate load forecasting is difficult for electricity operation because it depends on the historical load data and other influential external factors such as working days, seasonal, calendar, etc. As a result, the efficient classification model could improve forecasting accuracy. This research proposes combining classification and regression tree (CART) with pruning conditions and a deep belief network (DBN) forecasting model to provide an efficient electricity forecasting system for electricity generation. Firstly, the CART is built to classify the load data (48 periods per day) provided by the Electricity Generating Authority of Thailand (EGAT). The independent variables of CART are the day of the week, the month of the year, holiday or nonholiday, and bridging or nonbridging holiday. After creating the original CART, it is pruned with pre-defined conditions, and then the input load data are classified into the tree s leaf nodes. DBN models are then trained for every leaf node using these classified input data. Lastly, each testing set is provided and falls into the associated trained DBN model to measure the forecasting efficiency. The proposed model is compared with other deep learning (DL) models such as deep neural network (DNN) and long short-term memory (LSTM) that use the manual classification (MC) based on the day of the week in the fifth experiment. The other four experiments are also conducted by applying multiple forecasting models, different training and testing sets, different input structures, different data cleaning techniques. Machine learning models such as artificial neural networks (ANN) and support vector machines (SVM) are additionally implemented to compare forecasted results. Mean absolute percentage error (MAPE) is commonly used as an error measurement for all experiments. All prop