From Load to Net Energy Forecasting: Short-Term Residential Forecasting for the Blend of Load and PV Behind the Meter

Abstract

As distribution networks worldwide are experiencing the adoption of residential solar photovoltaic (PV) more than ever, the need for transiting from the concept of load forecasting to net energy forecasting, i.e. predicting the blend of PV and load as a whole, is pressing. While most of the existing literature has focused on load forecasting, this paper, for the first time, contributes to this transition at both single household and low aggregate levels through a comprehensive study. The paper also proposes a multi-input single-output (MISO) model based on an efficient long short-term memory (LSTM) neural network, by which different household energy profiles help provide more accurate forecasts for other households or aggregate energy profile. This technique, indeed, considers the spatial dependencies of households' profile indirectly. Through this study, the underlying problem of short-term net energy forecasting is compared to load forecasting, and it is shown how the inclusion of PV generation behind the meter could deteriorate forecasting accuracy. Moreover, the impact of the level of granularity associated with smart meter data on the aggregated net energy forecasting is discussed, and it is revealed that the higher resolution data can potentially alleviate the accuracy lost. Furthermore, online LSTM, as opposed to proposed batch learning MISO LSTM, is used as a forecasting tool. The results show online LSTM is more resilient to sudden changes at the single household level, while MISO LSTM is efficient for aggregate level. The proposed framework is conducted on two real Ausgrid and Solar Analytics case studies in Australia.

Keywords

• Deep learning
• long short-term memory (LSTM)
• recurrent neural networks
• residential load forecasting
• short-term net energy forecasting
• smart meter