Multi-Scale Solar Forecasting Using Wavelet Decomposition

For effective grid integration, accurate intra-hour to hour-ahead energy forecasting is critical.  Currently, simple persistence methods, in which energy production is considered static, are the industry standard for short-term forecasting.  Though easy to implement and more accurate than numerical weather prediction, persistence methods fail under dynamic weather conditions.  For instance, utility-scale solar plants can be subjected to ramp rates in excess of 5% capacity per minute.  Under these conditions, a persistence model will fail to react to the production variability and short-term forecast accuracy will be poor.  To improve upon this, particularly for variable conditions, DNV GL has developed a multi-scale artificial neural network (ANN) model for solar power forecasting.  

Typically, ANN models are built using a time-series history of observed power data.  Even without exogenous inputs such as numerical weather prediction data, ANN models have been shown to outperform persistence for forecast horizons of less than several hours.  Generally, observed data is temporally averaged prior to model creation.  Though this can improve mean-error statistics, it eliminates intra-hour variability.  Subsequently, the ANN model cannot predict short-term ramp events.  Conversely, ANN forecasts built using non-temporally averaged data are often noisy and fail to capture hourly or longer trends in energy production.  Fundamentally, this is because different time-scales have different levels of predictability.  In order to preserve intra-hour variability while minimizing the mean-error statistics, this study implements an ANN model which is built upon multiple temporal scales of observed data.

To design this model, observed high-resolution power time-series are first partitioned by temporal scale using wavelet decomposition.  Wavelet decomposition is a method of separating variability at different time-scales from an underlying temporally-averaged signal.    For each time-scale, independent ANN models are constructed and optimized using a genetic algorithm.  Next, the ANN models are applied to an independent evaluation period to create temporally-averaged and variability forecasts for forecast horizons of 5 minutes to 3 hours.  Finally, the forecast variability from each time-scale is added to the forecast average time-series to create the final forecast.  In this study, this is applied to 6 months of anonymized production data from a utility-scale solar plant.  Overall, this forecast will be shown to improve over both persistence and auto-regressive methods, particularly under dynamic weather conditions.