A robust neural network filter for electricity demand prediction

Abstract

This paper is concerned with one-day-ahead hourly predictions of electricity demand for Puget Power, a local electricity utility for the Seattle area. Standard modelling techniques, including neural networks, will fail 'when the assumptions of the model are violated. It is demonstrated that typical modelling assumptions such as no outliers or level shifts are incorrect for electric power demand time series. A filter which removes or lessens the significance of outliers and level shifts is demonstrated. This filter produces 'clean data' which is used as the basis for future robust predictions. The robust predictions are shown to be better than non-robust counterparts on electricity load data. The outliers identified by the filter are shown to correspond with suspicious data. Finally, the estimated level shifts are in agreement with the belief that load growth is taking place year to year.

Neural networks for short-term load forecasting: a review and evaluation

“…The first way was by repeatedly forecasting one hourly load at a time, as in [28], [29]. The second way was by using a system with 24 NNs in parallel, one for each hour of the day: [61] compared the results of such a system to those of a set of 24 regression equations; [18] considered the load as an stationary process with level changes and outliers, filtered these out by a Kalman filter, and modeled the remaining load by a NN system; [86] considered the load as the output of a dynamic system, and modeled it by a set of 24 recurrent NNs.…”

“…Since the load series are often nonstationary, [17] suggested that NNs could be used to model the first differences of the series, as nonlinear extensions to the ARIMA models. Other authors dealt with the problem of nonstationarity by detrending the series [63], [83], [84] or by filtering it with a Kalman filter [18].…”

“…Pre-processing may also be needed to "clean" the data, by removing outliers, missing values or any irregularities, since NNs are sensitive to such defective data. References [52], [72] devised heuristics to regularize their data; [18] filtered out the irregularities with a Kalman filter.…”

Using combined forecasts with changing weights for electricity demand profiling

“…Reassuringly, this is the case in expression (7). The definition of the dummy variable W t implies that if it takes a value of 1, the weight on f t,1 should be greater than if W t takes a value of 0.…”

“…However, there is no consensus as to the best approach to short-term electricity demand forecasting. Bunn 5 reviews a wide variety of methods and the recent competition organised by the Puget Power Company in Seattle witnessed a range of different approaches including: time-varying splines 6 , artificial neural networks 7 , multiple regression models 8 , judgemental forecasts produced by Puget Power's own personnel, and BoxJenkins transfer function intervention-noise models. NGC's cardinal point forecasts are produced by separate regression models which are functions of seasonal and weather variables.…”