Towards An Enhanced Backpropagation Network for Short-Term Load Demand Forecasting

Artificial neural networks (ANNs) are ideal for the prediction and classification of non-linear relationships however they are also known for computational intensity and long training times especially when large data sets are used. A two-tiered approach combining data mining algorithms is proposed in order to enhance an artificial neural network’s performance when applied to a phenomenon exhibits predictable changes every calendar year such as that of electrical load demand. This approach is simulated using the French zonal load data for 2016 and 2017. The first tier performs clustering into seasons and classification into day-types. The second tier uses artificial neural networks to forecast 24-hour loads. The first tier results are the focus of this. The K-means algorithm is first applied to the morning slope feature of the data set and a comparison is then made between the Naïve Bayes algorithm and the k-Nearest Neighbors algorithm to determine the better classifier for this particular data set. The first tier results show that calendar-based clustering does not accurately reflect electrical load behavior. The results also show that k-Nearest Neighbors is the better classifier for this particular data set. It is expected that by optimizing the data set and reducing training time, the learning performance of ANN-based short-term load demand forecasting.