A Short-Term Household Load Forecasting Framework Using LSTM and Data Preparation

Abstract

IoT devices are deployed in a building to instantly collect electricity load usage for next hour load consumption forecasting so that the operation of the building can be properly managed. However, due to the hardware or system error, data missing or overflow might occur. Noise might also be added into the collected data. Under this circumstance, the accuracy of next hour load forecasting would degrade which in turn decreases the building operation performance. In this paper, we propose an hourly load forecasting framework combining Data Preparation and LSTM, namely LSTM-DP, by considering data pre-processing, feature engineering and Long Short-Term Memory (LSTM). In LSTM-DP, the collected data is firstly processed by interpolation and Savitzky Golay filter, therefore the pattern of load consumption can be properly extracted by LSTM for next hour load forecasting. Moreover, we adopt two-stack LSTM to better determine the relationships among the time series information. We study the real data collected from three buildings of a company in Asia to investigate the performance of next hour load forecasting, and the results show the proposed LSTM-DP outperforms others.

Keywords

• Short-term load forecasting
• long short-term memory (LSTM)
• smart meter
• data pre-processing
• feature engineering