Modeling Regional Electricity Load in India

Abstract

Electricity as a product cannot generally be stored. Hence, it is required to match demand and supply on a real time basis in order to avoid disturbance in grid frequency and consequently ensure quality of power supply. The agencies involved in scheduling of power in India divide a day into ninety-six time buckets - each bucket of fifteen minutes duration. The load is matched for each time bucket. This matching is done the night before the start of actual dispatch. In fact, the demand supply schedule for a day is finalized at 11 PM on the previous day. The regulated bulk supply tariff in India has an unscheduled interchange component, which is linked to grid frequency. Thus, any smart generator of electricity would make abnormal profits provided she is able to forecast the load properly. The present paper attempts to model the hourly load in the northern grid in India. The hourly load is estimated by aggregating load over five-minute intervals. Data corresponding to each hour is treated as a single time series and each series is modeled independently. The paper has used an estimation window of eleven months and forecast window of one month. Autoregressive models with dummy variables to capture (a) dayof- the-week effect, (b) national non-Sunday holiday effect, and (c) seasonality effect turn out to be quite effective in explaining the hourly load behavior. The results show evidences of clustering of load behavior. Five clusters of time period within a day were observed where the load behavior can be captured with a single model. These clusters take care of sixteen hours of a day. Each of the remaining eight hours behaves differently. Interestingly, the behavior of load around mid night and during morning hours of the day does not depend on day of the week. Furthermore, Thursdays and Saturdays had the least impact on the hourly load. Another interesting finding is that there is no effect of national non-Sunday holidays on the load. The mean absolute percentage error of the best-fit model in calculating one-day-ahead out-of-sample logload forecast is quite small (ranging from 0.14% to 0.20%). The findings of the paper may have profound implications for the regulator - (a) the regulator can use the information to introduce time-of-the-day pricing; (b) a fifteenminute time bucket may be desirable for load scheduling, but for load forecasting one may use the clusters. Thus, the electricity generator may schedule its generation based on cluster-wise load forecast, rather than hourly forecast.