Improving duck curve by dynamic pricing and battery scheduling based on a deep reinforcement learning approach

Duck Curve Aware Dynamic Pricing and Battery Scheduling Strategy Using Reinforcement Learning

“…This modeling helps mimic a real-world system as a proof-of-concept for the proposed method. This paper extends the work presented in [40]. The earlier work only focused on load curtailment in response to retail prices as prosumer behavior.…”

“…Without future information and the direct control of the prosumer, the performance of DRL-avg is close to Optimal. Furthermore, DRL-avg also outperforms NoShift [40] for both metrics. DRL-avg achieves a reduction in both the standard deviation and the PAR compared to other baselines.…”

“…• NoShift [40]: DRL-based strategy proposed in the previous work [40]; the prosumer has no load shifting (load deferral) mechanism (only reducing demand according to retail prices) and always charge/discharge the battery at maximum rates as described in Sections II-A-1 and II-A-3, respectively. The MDP model do not include important states, Σ t−1 k=1 e def er k,n and E dev t , as explained in Section III-B.…”

“…While the work in [40] considered the load curtailment, we specifically focus on load deferral, i.e., load shifting to later time slots, to make the prosumer behavior more realistic. In load deferral, the reduced load due to high retail electricity prices, e.g., stopping the operation of washing machines, does not disappear and will be reactivated in a low-price period or after a certain period [42].…”

“…Note that the previous study [40] assumed that the prosumer would always charge or discharge the battery at maximum rates whenever retail prices were either below or above a predetermined threshold λ th n . However, this controller did not allow for continuous control of the battery and did not consider current demand and PV generation.…”