PRT products and services are based on adaptive machine learning and intelligent systems concepts. We utilize multiple AI technologies such as artificial neural networks, fuzzy logic, genetic algorithms, and evolutionary computing. The common factor among these technologies is their ability to use historical data to capture and learn the complex and nonlinear relationships that exist between the forecasted variables and the factors that impact them. Each of these techniques uses a different approach for doing this task resulting in variations that are complementary. We utilize this attribute by blending multiple model forecasts through an adaptive blending process that “learns” how to optimally combine them for various conditions. This “Combination of Experts” approach yields a final forecast that is more accurate than any single forecasting technique.
Another unique aspect of our technology involves the use of an ongoing and continual adjustment and calibration of the models after they have been trained using historical data. This ongoing calibration leverages the most recent actual observed load, weather, price, generation and other relevant data. When new information becomes available, our models are update and a revised forecast is produced automatically. These continuous adjustments improve tracking accuracy and enable quick adaptation to situations that have not been present in the historical training data.
Weather has a major impact on the variables we forecast. We use a unique “Combination of Experts” approach utilizing forecasts from four major weather service providers to produce the most accurate weather forecasts. These different forecasts are combined based on their accuracy records and other relevant criteria. This approach yields a higher level accuracy than any individual vendor can provide.
Our forecasts are updated continuously on an hourly/sub-hourly basis using the most recent observed weather (or updated weather forecast), load, price, generation and other relevant data. For power markets, actual data is extracted from market’s web site as soon as it becomes available. This updating also performs a “forward error correction” process to remove possible biases.
We have an extensive quality control system in place to detect and fix data uploaded by user or extracted from market web sites. Our experienced technical team continually monitors the performance of forecast models. If a degradation of accuracy is detected, the models are recalibrated and retrained immediately to achieve optimal performance.