A Framework to Forecast Electricity Consumption of Meters using Automated Ranking and Data Preprocessing

Abstract

Forecasting electricity consumption is crucial for the operation planning of distribution companies and suppliers and for the success of deregulated electricity markets as a whole. Distribution companies often need consumption forecasting for meters to better plan operations and demand fulfillment. Although it is easier to forecast the aggregated demand for a region, meter based demand forecasting brings challenging issues such as non-uniform usage and uncertain customer consumption patterns. The stochastic nature of the demand for electricity, along with parameters such as temperature, humidity, and work habits, eventually causes deviations from the expected demand. In this paper, real meter data from a regional distribution company is used to cluster the customer using their non-uniform usage and automated ranking mechanism is proposed to select the best method to forecast the consumption. The proposed end-to-end methodology includes data processing, missing value detection and filling, abnormal value detection, and mass reading for meters and is applied to regional data for the period 2017-2018 and provides a powerful tool to forecasts the demand in hourly and daily horizons using only the past demand data. Besides proposing effective methodologies for data preprocessing, 10 different regression methods, 7 regressors, 5 machine learning methods that include LSTM and Ar-net models are used to forecast the meter based consumption. The hourly forecasting errors in the demand, in the Mean Absolute Percentage Error (MAPE) norm, are less than 4% for most customer groups. The meter based forecast is then aggregated to reach a final demand which is then used for operation and demand planning. The proposed framework can be considered reliable and practical in the circumstances needed to make demand and operation decisions.