Day-Ahead Prediction of Distributed Regional-Scale Photovoltaic Power

Day-ahead forecasts are required by electricity market investors to make informed decisions on the trading floor. Whereas it is relatively easier to predict the performance of a few large-scale photovoltaic (PV) systems, a large number of small-scale PV systems with a wide geographical spread poses more challenges because they are often not metered for real-time monitoring. This paper proposes an artificial neural network (ANN)-based model to achieve regional-scale day-ahead PV power forecasts based on weather variables from numerical weather predictions (excluding solar irradiance) as inputs. The model was first implemented by dividing a region into clusters and selecting a representative site for each cluster using data dimension reduction algorithms. Solar irradiance forecasts were then generated for each representative PV system and the corresponding PV power was simulated. The cluster power output was obtained using a linear upscaling model and summed to produce regional-scale power forecasts. The model's accuracy is validated using power generation data of several distributed systems in California. The results show at least a 29-percent root mean square error reduction over the benchmarking models.

Automated summary

This paper proposes an artificial neural network (ANN)-based model for regional-scale day-ahead PV power forecasts using weather variables from numerical weather predictions as inputs. The model divides a region into clusters, selects a representative site for each cluster, and generates solar irradiance forecasts and corresponding PV power simulations. The cluster power output is obtained using a linear upscaling model and summed to produce regional-scale power forecasts. The accuracy of the model is validated using power generation data of distributed systems in California, showing a 29% reduction in root mean square error compared to benchmarking models.