Data‐driven forecasting of local PV generation for stochastic PV‐battery system management. (9th May 2021)

• Record Type: Journal Article
• Authors: Kaffash, Mahtab; Bruninx, Kenneth; Deconinck, Geert
• Is Part Of: International journal of energy research. Volume 45:Number 11(2021)
• Page Start: 15962
• Page End: 15979
• Publication Date: 2021-05-09
• Subjects: data‐driven forecasting -- nonparametric distribution -- PV‐battery system -- stochastic optimization -- uncertainty of PV system; Power resources -- Periodicals; Power (Mechanics) -- Periodicals; Power resources -- Research -- Periodicals; 621.042
• DOI: 10.1002/er.6826
• Languages: English
• ISSNs: 0363-907X

Abstract

Summary: Power systems face more uncertainty by increasing photovoltaic system installations on the roof of buildings. To optimally manage energy and available flexibility in a building, stochastic optimization is used to take an optimal decision under uncertainty and minimize the operational cost. In stochastic optimization, a scenario set is used as an input to represent the uncertainty in a random variable, PV energy generation in this case. In this paper, a data‐driven method is proposed to obtain the distribution of the random variable and later generate scenario sets representing the uncertainty on day‐ahead PV energy generation installed on the roof of a building. This method is only based on historical PV generation and it does not require any other external data such as weather forecasts. A machine learning‐based technique is applied to forecast the PV energy production following by generating a scenario set for day‐ahead decision‐making. Later, the day‐ahead PV‐battery system management problem is formulated as a two‐stage stochastic optimization while the generated scenario set is the input of this optimization. The proposed algorithm is tested in the day‐ahead scheduling of a PV‐battery system for a commercial building, informed by real‐life measurement data. The results show that the proposed algorithm is able to capture the uncertainty in PV system while providing a cost‐optimal and reliable solution to the application problem. Without using any weather data, the error of the proposed PV energy forecast method reaches to NRMSE = 11.89, while there is 5% reduction in the operational cost of the proposed two‐stage stochastic optimization. Moreover, the proposed algorithm is easy to implement in the energy management system of a building to manage PV‐battery system. Abstract : A machine learning learning based point forecast followed by a nonparametric approach is proposed to obtain a predictive distribution and generate scenarios to represent the uncertainty of PV production in a building. The proposed method is relied on only historical local PV data, without using any external data set (weather and PV forecast data). The proposed approach is simple but effective, and can be implemented in buildings' energy management system to manage flexibility and energy while taking into account the uncertainty of local PV generation.