Has Wind Energy Forecasting Solved the Challenge Posed by Intermittency?: Evidence from Great Britain

A growing literature has affirmed with great confidence the feasibility of high penetration levels of wind energy in electric generation under existing technology.  This optimism is based in part on the belief that improvements in wind energy forecasting have effectively “solved” the challenges posed by wind energy’s intermittency.  For example, the UK’s Royal Academy of Engineering (2014, p. 33) has noted that wind energy’s capacity weighted forecast error of about five percent is evidence that the wind energy forecasts in Great Britain are highly accurate.  In its words, “Although wind energy cannot be dispatched by the system operator, its level of output can be predicted to a high degree of accuracy.” Using data from Great Britain, this paper examines this claim.  

Following Forbes et. al (2012), the analysis begins with an assessment of the accuracy of day-ahead wind energy and day-ahead load forecasts over the period 1 January 2012 through 31 December 2013.  The findings indicate that the energy weighted root-mean-squared error (EWRMSE) of the day-ahead wind energy forecasts is considerably larger than the EWRMSE of the day-ahead load forecasts.  Evidence is also presented that the capacity weighted root-mean-squared error (CWRMSE) of the day-ahead wind energy forecasts is considerably larger than the CWRMSE of the day-ahead load forecasts.  The analysis also  indicates that while the day-ahead load forecasts are more accurate than a persistence load forecast, the day-ahead wind energy forecasts are less accurate than a persistence wind energy forecast.  Specifically, with a persistence forecast as a reference, the mean-square-error-skill-score (MSESS) of the day-ahead load forecasts is positive while the MSESS of the day-ahead wind energy forecasts is negative.  This result is not unique to Great Britain.  

In Great Britain, the final physical notifications of expected generation submitted by the wind energy generators to the system operator one hour prior to real time are a near real-time predictor of the actual  wind energy outturn.  Unfortunately, this near real-time predictor of wind energy generation is highly inaccurate relative to the notifications provided to the system operator by the generating stations fueled by coal, natural gas, or nuclear energy.  This is corroborated by the energy imbalances by fuel type (Figure 1).  Observe that wind energy’s average energy imbalance relative to its corresponding level of metered generation is much higher than for other energy sources.  This occurs because the level of scheduled generation declared by the wind energy producers, i.e. the final physical notification, is a highly inaccurate predictor of actual wind energy production.  Moreover, the actual level of wind energy generation is systematically less than than the scheduled level of wind energy generation even if  wind energy curtailments are factored into the  analysis (Figure 2). One implication of this finding is that a substantial portion of the intended wind energy production is actually being accounted for by fossil fuel generation, fossil fuels being  the primary energy source that the system operator relies on resolve energy imbalances.  Another implication is that higher levels of wind energy pentration may have adverse implications for the reliability of the electric power system in Great Britain.

With the above issues in mind, the paper examines whether an improvement in the accuracy of the wind energy forecast is possible.  Our starting point is a least squares regression of the day-ahead wind energy forecast  errors on the day-ahead forecasted weather conditions for a location in Scotland that yields highly statistically significant coefficients on the following variables: forecasted temperature, forecasted wind speed, forecasted humidity, forecasted dewpoint, forecasted visibility, and the forecasted probability of precipitation. In short,  the wind energy forecasts do not fully reflect the information contained in the day-ahead weather forecasts. Moreover, the results  indicate that the wind energy forecast errors have the undesirable property of being correlated with the forecasted level of wind energy.  Time series analysis also indicates that the wind energy forecast errors do not have the property of “white noise.”     With these insights, we estimated an econometric model using half-hour data over the period 1 January 2012 - 31 December 2013.  Preliminary evidence suggests that a substantial improvement in near real-time forecast accuracy is possible by modifying the forecast information based on the time-series model.  Specifically, the out-of-sample RMSE of the time-series augmented forecast is less than half the RMSE of the existing near real-time projection over the same time period and is less than one third the RMSE of the day-ahead forecast.  These methodological findings may mitigate the challenges posed by wind energy’s intermittency.