 |
Center for Nonlinear Studies |
For wind power producers (WPPs) who bid in day-ahead (DA) electricity markets, forecast errors can cause large penalties in the real-time (RT) market if the WPP defaults on its cleared DA offer. Battery energy storage systems (BESSs) can be strategically charged and discharged to reduce the impact of wind power forecast errors. This talk will discuss a novel hybrid stochastic optimization (SO) with twin delayed deep deterministic policy gradient (SO-TD3) algorithm for co-optimizing wind offers to the DA/RT electricity markets and the operation strategy of an on-site BESS. First, the DA wind offer curves and an initial BESS operation strategy are found using SO. Uncertainty in wind generation and market prices is considered through scenario generation. Then, for each price/power point on the offer curve, a TD3 agent is designed to re-optimize the BESS operation strategy. The TD3 algorithm uses several modifications to ensure safe BESS operation, and the likelihood of default on the DA wind offers is modeled and incorporated into the agent's strategy. A custom reward function is designed to improve profit compared to the SO-only optimization, while keeping the BESS SOC within an acceptable range. The proposed method is compared to optimizing both the wind offers and BESS strategy using SO. Case study results show that over 100 test days, the SO-TD3 method improves the net profit of the wind-BESS system by 0.51% and 2.45% in risk-averse and risk-neutral cases, respectively. The proposed method also reduces the number of hours with extremely low BESS state-of-charge from 365 to 24 in the risk-averse case, and from 366 to 27 in the risk-neutral case, which can increase the battery lifetime. In addition, SO models for optimizing bidding strategies for a WPP that participates in US energy and reserve markets will be discussed. Finally, this talk will discuss a two-stage machine learning-based framework for net load forecasting in areas with limited observability and high behind-the-meter (BTM) PV generation. First, the measured net load data is disaggregated into separate profiles of PV and pure load. LSTM models are used to forecast the PV generation and pure load individually, and the results are combined for a net load forecast. A compensator is also designed to correct the error of the net load forecast, using historical forecast errors of the PV generation, pure load, and net load. Case study results show that the proposed framework reduces the net load forecast error compared to benchmark models, and that when the net load forecast error is periodic, the compensator can correct the error to improve the forecast accuracy.
Host: Anatoly Zlotnik (T-5)