Using Deep Learning Techniques for South African Power Distribution Networks Load Forecasting

Load forecasting has many benefits for utilities. Artificial Intelligence (AI) has been seen to be effective in load forecasting. Deep Learning AI techniques have been found to perform better than traditional AI techniques. The study of deep learning techniques application in South African load forecasting is in its infancy. This paper presents a study of a South African distribution substation using two deep learning techniques, deep belief networks (DBN) and long short-term memory (LSTM), to address this. The impact of temperature and cleaning up the loading data is also studied. It was found that an LSTM model achieved the lowest errors with a symmetric mean absolute percentage error (sMAPE) of 3.3%, an), mean absolute error (MAE) of 4.6% and root mean square error (RMSE) of 5.5%. These errors were achieved with non-cleaned data with temperature not used as a variable for the training the model. Deep learning techniques can thus be used without weather parameters to forecast distribution substation data with low errors.