The Effect of Data Length on Holt-Winters and LSTM in Knowledge Discovery for Electricity Consumption Forecasting

Abstract

Electricity consumption forecasting is an application of knowledge discovery influenced by the length of historical data, even though adding more data does not always improve performance and can instead increase computational burden. This study analyzes the effect of varying training data lengths (1, 3, 5, 7, and 9 years) on forecasting accuracy by comparing Holt-Winters and Long Short-Term Memory (LSTM) on the daily PJME electricity consumption dataset, using a fixed 90-day test set, evaluated through MAE, RMSE, MAPE, and computation time. Results show that Holt-Winters achieves its optimal point at nine years of training data (MAPE 9.37%), while LSTM is optimal at three years (MAPE 9.69%) but requires 51 times longer computation time. This crossover pattern indicates that Holt-Winters performance is determined by the representativeness of the initial seasonal component estimate, whereas LSTM is affected by error accumulation in the recursive multi-step forecasting scheme, confirming that the required length of historical data depends on the method's characteristics. The contribution of this study is a fair comparative experimental design that isolates the effect of training data length from the effect of the forecasting horizon, a design not previously applied to regional-scale electricity consumption forecasting, thereby filling a methodological gap in similar studies. Practically, these findings provide concrete guidance for energy providers in determining historical data management strategies Holt-Winters is recommended for providers with long historical data but limited computational infrastructure, while LSTM is more suitable for providers with limited historical data but adequate computational resources, allowing data and computational investments to be directed more efficiently.