Scheduling More Wind: Exploiting Probabilistic Forecasts

Every day at wind plants across the U.S. there is an important daily task that is approached as a matter of routine. That task is providing the required schedule of tomorrow's expected wind output for the day-ahead electricity market. Typically the professionals carrying out this task are risk adverse and often use very simple rules of thumb to submit their day-ahead schedules. One such example is submitting only half of the expected output to minimize the risk of getting caught short in the real-time market with the view that everything will “settle out” in the real-time market. Our experience suggests that this is a major lost opportunity, possibly by as much as 3 giga-watt hours (GWh) for an average 100 mega-watt (MW) plant each month. While many wind plant energy schedulers agree that there is more money to be made with a risk-adjusted scheduling strategy for the day-ahead market, they often lack the tools and information required to extract additional value from this daily routine. The solution is a better understanding and exploitation of calibrated probabilistic forecasts, which allow more energy to be scheduled into the day-ahead market without increasing the risk of under delivering that energy.

To demonstrate the benefits of applying a probabilistic approach to wind energy scheduling, we have conducted a 6-month study of day-ahead forecasting comparing the results of using a P-value versus a deterministic forecast “haircutting” approach. To compare the techniques fairly, we chose the closest P-value that matched the downside risk level of "haircut" strategy, which was close to 30%. Therefore we selected the P70 forecast value. While the downside risk of the two approaches is nearly equal, the value of the probabilistic approach becomes apparent when one compares the amount of energy scheduled by both strategies. Over the 6-month period from May-October 2013, the amount of energy scheduled by the “haircutting” method is 62.1 GWh while the P70 achieves a total of 82.6 GWh, resulting in a net gain of 20.5 GWh. Even though both approaches had nearly equal downside risk profiles, the probabilistic approach was superior in scheduling more energy.