Improved short-term load forecasting using bagged neural networks

• We present a load forecasting scheme using bagged neural networks that improves load forecasting accuracy.
• The bagged neural networks consist of creating multiple sets of the same data by sampling randomly with replacement, training a neural network on each data set, followed by averaging the results.
• The bagged neural networks reduce load forecasting error and variation in the estimation accuracy that occurs if a single artificial neural network is used.
• We use data from the New England Pool region, compare the results with several existing techniques and show the improvement offered by our method.

In this paper we present improved short-term load forecasting using bagged neural networks (BNNs). The BNNs consist of creating multiple sets of data by sampling randomly with replacement, training a neural network on each data set, and averaging the results obtained from each trained neural network. The bagging process reduces estimation errors and variation range of errors compared to using a single neural network for load forecasting. Examples with real data show the effectiveness of our proposed techniques by demonstrating that using BNNs can reduce load forecasting errors, compared to various existing techniques.