Application of Self-Organized Map Informed Artificial Neural Networks for Load and Wind Forecasting

The accuracy of wind forecasts is becoming increasingly essential as grid operators continue to integrate additional renewable generation onto the electric grid. Forecast errors affect rate payers, grid operators, wind plant maintenance crews and energy traders through increases in prices, project down time or lost revenue. While extensive and beneficial efforts were undertaken in recent years to improve statistical and physical weather models for a broad spectrum of applications these changes have often not been adopted by the renewables industry either due to insufficient forecast improvements, budget constraints, insufficient computational resources to adapt the new technique to hourly or 5 minute forecast updates, changes in customer requirements or simply an unawareness of these new techniques by industry forecasts vendors.

DNV GL presents a real world application of a novel statistical forecasting technique in which an automated tool, based on a combination of Artificial Neural Networks (ANN) and Self-organizing Maps (SOM), has been developed to identify different weather regimes as they impact forecast accuracy and train specific statistical models for each regime. This work, based in part on a study by McCandless et al, demonstrates the value of these enhanced operational forecasts at actual wind plants through reduced error and decision support assistance. In addition, DNV GL illustrates the flexibility of this technique and shows how it can also be applied to electrical load forecasting for a major utility.

Forecast accuracy can be dependent on specific weather regimes for a given location. To account for these dependencies it is important that parameterizations used in statistical models change as the regime changes. Kohonen Self Organizing maps are used to identify relationships between weather regimes at model initialization time and day-ahead forecast error. Weather regime dependent model combinations and parameterizations are then determined and used to create a new optimized forecast for the specific weather regime on an hourly basis. For this study ANN models are trained to combine multiple Numerical Weather Prediction (NWP) data sets for each specific regime. By identifying the weather regime at model initialization time the tool is able to choose the model combination or configuration that performs best for a given set of measured or initial conditions. The creation of distinct models for each large scale weather pattern across a region allows the artificial neural networks to be optimized by effectively reducing unnecessary data that may cloud the solution and making it easier for each neural network to converge on an accurate solution.

Although the use of clustering techniques to identify specific and common weather regimes is not new and has been used in a number of climate and weather pattern studies it has seen little to no use for relating forecast error to atmospheric conditions and has likely never been used in an operational forecast service for wind generation or load.