Irradiance to Power Conversion Model Creation
Methodology
The solar power forecasting system forecasts power generation in two stages. The first stage involves forecasting irradiance such as global horizontal irradiance. The second stage involves converting the forecasted irradiance to power. In order to convert the forecasted irradiance to power, empirical data consisting of irradiance, power pairs are gathered over a sufficiently long period of time for an individual solar farm. These data pairs are then utilized to create an irradiance to power conversion model for that farm.
Data
The data needed to create an irradiance to power conversion model consists of observations of irradiance, either global horizontal irradiance (GHI), direct normal irradiance (DNI) or plane of array (POA), matched with observations of power. When matching the observations we chose to average the values over a fixed time interval. This ensured observations were matched to a common time and eliminated issues related to differing data update rates and mismatched observations. It is best to have the observations on the same time scale as the forecast, i.e. for forecasts every 15 minutes matching observations of irradiance and power every 15 minutes is ideal. For the majority of the models we created this was possible. However, for a few farms this was not the case and observations were matched on an hourly basis. Data was collected for the various farms over periods of one year up to 3 years, with a longer history of data producing a more robust model.
The model training data set consisted of matched irradiance and power observations along with information about the hour of day, day and week of the year for the data as well as values for the solar azimuth and solar elevation for the observation location and time.
Quality Control
Quality control issues with the data are prevalent and quality controlling the training data set consisting of matched irradiance and power observations is critical. Obvious issues were cases where one of the values from either the irradiance or power would be stuck on a specific value while the other fields continued to vary. Another frequent occurrence would be power values of 0 with varying irradiance, and vis-versa. Plotting the matched irradiance and power values illuminated these issues but also displayed some values falling outside the standard envelope of observations for each specific farm. Stuck values along with any values outside the standard envelope of observations for each particular farm were eliminated from the training set. Plotting power verses solar elevation also showed some observations with large power values but below zero solar elevation. These few observations were also eliminated.
Model Creation
The irradiance to power conversion models were created using the rule based regression tree software package Cubist by RuleQuest. Cubist was chosen due to the previous success of using the software to create wind to power conversion models for another application, along with familiarity and ease of use. A different model was created for each farm, with some models using different predictor variables. In general, irradiance was always used along with variables that identified time and season. Some combination of hour of day, day and week of year, solar elevation and azimuth were used in each model. Our models also took advantage of the Cubist option to create a committee model, which allows for a group of several rule based models to be created that each predict the target value, with the final answer being the average of all the rule based models predictions. Each subsequent committee member model after the initial rule based model attempts to correct the errors from the previous model. The committee models appeared to give a smoother forecast prediction through time when compared to a basic single rule-based model. Other options, such as using a nearest neighbor from the training set to modify the prediction, were considered but did not seem to improve the final forecast and were discarded.