US 11,808,915 B2
Wind power prediction method and system based on deep deterministic policy gradient algorithm
Ming Yang, Jinan (CN); Menglin Li, Jinan (CN); Yixiao Yu, Jinan (CN); and Peng Li, Jinan (CN)
Assigned to SHANDONG UNIVERSITY, Jinan (CN)
Filed by SHANDONG UNIVERSITY, Jinan (CN)
Filed on Mar. 9, 2023, as Appl. No. 18/119,450.
Claims priority of application No. 202210229898.4 (CN), filed on Mar. 10, 2022.
Prior Publication US 2023/0288607 A1, Sep. 14, 2023
Int. Cl. G01W 1/10 (2006.01); G05B 13/02 (2006.01); G06N 3/092 (2023.01); G06N 3/047 (2023.01)
CPC G01W 1/10 (2013.01) [G05B 13/026 (2013.01); G06N 3/092 (2023.01); G06N 3/047 (2023.01)] 7 Claims
OG exemplary drawing
 
1. A wind power prediction method based on a deep deterministic policy gradient (DDPG) algorithm, comprising:
obtaining data related to a wind power prediction;
inputting the obtained data to each of a plurality of different trained prediction sub-models, to obtain corresponding wind power prediction value of each prediction sub-model;
building a combined model, wherein the combined model is a combination of the plurality of prediction sub-models, and each prediction sub-model is assigned with a respective weight;
perceiving a current state from a prediction environment at a to-be-predicted time point by using the DDPG algorithm, determining a policy based on the current state, obtaining a weight with exploration noise, assigning the weight to the combined model, and iteratively optimizing the policy based on a feedback reward until the DDPG algorithm converges, wherein the DDPG algorithm comprises three basic elements: a state, an action, and a reward, the state is an indicator reflecting external environment information, comprising fluctuation information of meteorological prediction and prediction performance information of the prediction sub-models at a plurality of latest time points; the action is a determined weight and the reward comprises a fixed reward given based on a ranking, and an additional reward set based on a ratio of an absolute prediction error of the combined model to an absolute prediction error of an optimal prediction sub-model when the combined model ranks first; and
determining a final weight based on the converged current policy, and assigning the final weight to the combined model to obtain a final wind power prediction value.