Impact of Targeted Measurements and Advanced Machine Learning Techniques on 0−3-h Ahead Rapid Update Wind Power and Ramp Rate Forecasts in the Tehachapi Wind Resource Area of California

1. Introduction

The management of wind power variability is a key factor in minimizing the cost of integrating wind power into electric grid systems. Many North American system operators have indicated a need for more accurate very-short-term rapid update prediction of large changes in wind power production (i.e. “ramps”). The California Independent System Operator (CAISO) has indicated that the prediction of ramps in the Tehachapi Wind Resource Area (TWRA) would have great value due to the concentrated wind generation capacity there.

The California Energy Commission (CEC) and the Electric Power Research Institute (EPRI) supported a team lead by the University of California, Davis to improve the accuracy of TWRA wind forecasts. This paper presents results from the project component that focused on a time series prediction approach using targeted sensor data deployed for the project as well other onsite and offsite data as input into an advanced Machine Learning (ML) method. The objective was to produce forecast updates every 15 minutes for a look-ahead period of 0-3 hours. Numerical Weather Prediction (NWP) is NOT an effective tool for this update frequency and look-ahead period.

2. Approach

The approach was to deploy a targeted network of remote sensing devices in the TWRA region and use the sensor data (along with data from the wind generation facilities and other public sources) as input for a machine-learning-based prediction model.

The type and location of the deployed sensors are shown in Figure 1. The locations of the project sensors was based on a forecast sensitivity analysis [1] that used NWP-based Ensemble Sensitivity Analysis (ESA) to identify locations and types of measurement variables that would have the greatest beneficial impact on 0-3 hour TWRA power production forecasts. The suggested locations were adjusted due to practical deployment considerations.

Data from the targeted sensors and other sources were input into the Gradient Boosted Machine (GBM) machine learning method [2]. GBM employs a decision tree approach to modeling the relationships between the predictors and predictands. A new 15-minute resolution 0-3 hour forecast was generated every 15 minutes.

The GBM model was trained with a 24-month sample and evaluated on a 12-month independent sample. Two different predictands were considered for the TWRA aggregate: (1) the 15-minute average power production for each 15-minute interval in the 3-hour forecast period and (2) the maximum and minimum 60-minute ramp rate over the entire 3-hour forecast.

In order to assess the impact of the targeted sensor data, 5 experiments with different combinations of predictors were executed: (1) persistence of the power generation measured at the forecast issue time, (2) a GBM model with the persistence value from #1 and the time of day and day of the year, (3) a GBM model with predictors from #2 and the recent time series of data from the wind generation resources, (4) a GBM model with predictors from #3 publicly available offsite meteorological data, and (5) a GBM model with predictors from #4 plus the targeted network of sensors. Additional experiments were also conducted to determine the value of data from each of the remote sensing devices deployed in the project. Finally, experiments were conducted to determine the importance of customizing predictors by forecast look-ahead and season.

3. Results

Figure 2 depicts the mean absolute error (MAE) of forecasts from the “predictor” experiments. The red bars depict the forecasts that included the predictors from the project sensor data. There is very little impact on forecast performance at 15 and 30 minutes. However, the addition of project sensor data to data from all other sources reduced MAE for look-ahead periods of 90 minutes and longer by up to 10%.

Figure 3 shows the “sensor” experiment results. This chart illustrates the reduction in MAE when predictors from an individual sensor were added to all other predictors. Positive values indicate a reduction in MAE. Data from the Radar Wind Profiler (RWP) at the Bena site had the greatest impact. The Sodar at the Windmatic site which is immediately upstream of the wind production area had the second largest impact. The temperature profiling sensors had much less impact.

The results of the seasonal and forecast look-ahead customization forecasts will be presented at the conference.

4. Conclusions

The results from the 0-3 hour statistical prediction component of the Tehachapi wind forecast improvement project indicate that the deployment of a targeted array of remote sensing devices and the use of advanced machine-learning based prediction methods can reduce the MAE of 15-minute average power productions forecasts by 5% to 10% with respect to a forecast that uses data from all other available sources. Similar improvements were obtained in the large ramp rate event forecasts (i.e. a change of greater than 25% of capacity in 60 minutes). The CAISO has indicated that this level of improvement in ramp event forecasts can have significant benefits for managing wind variability in regions of highly concentrated wind generation resources.

References

[1] Zack, J, Natenberg, E., Young, S., Manobianco, J., Kamath, C., 2010: Application of Ensemble Sensitivity Analysis to Observation Targeting for Short-term Wind Speed Forecasting. Lawrence Livermore National Laboratory Report No LLNL-TR-424442,February 23, 2010, https://e-reports-ext.llnl.gov/pdf/387510.pdf

[2] Mason, L.; Baxter, J.; Bartlett, P. L.; Frean, Marcus, 1999: Boosting Algorithms as Gradient Descent. In S.A. Solla and T.K. Leen and K. Müller. Advances in Neural Information Processing Systems 12. MIT Press. pp. 512–518.