An improvement of photovoltaic power forecasting from numerical weather prediction

Abstract

This thesis presents forecasting models for improving the hourly predicted solar power from Numerical Weather Prediction (NWP) in one day horizon. A practical constraint is that the models must provide the predicted solar power before 14.00 hrs daily so that the Short-term Operation Planning Section, Generation Operation Planning Department, Power System Control and Operation Division, Electricity Generating Authority of Thailand can use the prediction to plan dispatching of the next day. Available data used in this research are from two sources: local weather measurements and power meter installed at the top of Electrical Engineering Building, Faculty of Engineering, Chulalongkorn University and NWP outputs. The prediction from NWP still has a significant error due to a complexity of cloud movement, so we propose models to refine the predicted solar irradiance from NWP. The models can be separated into two types according to time steps of updating rules: hourlystep and daily-step models. Each of these two models can also be divided into two groups according to mathematical equations of updating rules: model output statistics (MOS) which is a linear regression model and MOS used with Kalman filter (MOS+KF). This thesis has two contributions. Firstly, we applied statistical methods including partial correlation analysis, stepwise regression, and subset regression to select relevant inputs of MOS. The second contribution is that the limitation of data at the time of forecast was taken into consideration to modify the forecast updating rules by MOS+KF. The experimental results showed that MOS and MOS+KF models provide the root mean squared error of solar irradiace of 160 and 157 kW/m2, respectively, while the NWP error was 250 kW/m². When the irradiance forecasts were converted to the predicted solar power, the best among all proposed model was daily-step MOS+KF model which provided the normalized (by capacity of 8 kW) root mean squared error of 13 %, reduced from NWP by 9 %.