Competitive Neural Network Based Algorithm for Long Range Time Series Forecasting Case Study: Electric Load Forecasting

Time series forecasting takes the past values of a time series and uses them to forecast the future values. In this paper, we have proposed a new algorithm for multistep ahead time series forecasting. The original time series and differenced series are classified using competitive learning neural network. Transition matrix on the basis of transition from a class in original time series to the class of deformed series is formed. The last few values of the time series are used to find the best deformed series vector using transition matrix and hence future values of the time series are calculated as sum of test vector and differenced series vector. Long range forecasting is achieved by iterating the forecasted values of current iteration as the input for next iteration. The algorithm is validated for benchmark time series forecasting. We have also applied the algorithm to a real life problem of forecasting i.e. electric load consumption