Load forecasting based on hybrid models with artificial neural network

The primary purpose of power system planning is to meet customers' electrical energy needs economically, reliably and quality. One of the important steps in the planning process is load estimation. Traditionally statistical methods have widespread use because they are both easy to implement and perform well for load estimation. However, in recent years artificial intelligence methods have begun to be used extensively for estimating loads. In this study, artificial neural networks (ANN), adaptive neural fuzzy logic (ANFIS) and wavelet transform-artificial neural networks (WNN) were used to estimate the annual and seasonal loads for Eskişehir. Temperature, population, import, export and time data of Eskişehir were used as input data. Predicted methods' accuracy level was investigated. Trial models constructed using different ANN, ANFIS, WNN parameters were compared according to test MAPE values. The best performing YSA structure has 2 hidden layers, 12 neurons in hidden layers and uses tangent sigmoid transfer function and linear transfer function in the output layer. The best ANFIS structure is 64 rule grid partition method, trimf in the rule layer and linear transfer function in the output layer. The best DDYSA structure has single hidden layer with 12 neurons, using tangent sigmoid transfer function in this layer and using linear transfer function in the output layer. It has been observed that YSA has smaller MAPE values than other two methods in the annual and seasonal estimations of Eskişehir, thus giving better results.