ESS Control Scheduling Using Historical Reinforcement Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing energy storage system (ESS) scheduling control methods face challenges in accurately managing peak load due to limited data corresponding to the same season and time, especially when using reinforcement training models, which are not adequately trained on sufficient historical data.
Innovation Solution
A method that trains a reinforcement training model using past power consumption data to generate ESS control information, dividing it into weekday and holiday data, and applying it as daily ESS control information for present periods, ensuring strong responsiveness to instantaneous changes and reducing peak load by averaging historical data for similar seasons.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If reinforcement training model is trained using historical power consumption data, then ESS control performance is improved, but the amount of available training data is insufficient due to seasonal limitations
Solution Approach 1:
The patent applies preliminary action by training the reinforcement learning model in advance using historical power consumption data from previous years for the same seasonal period. This allows the model to be pre-trained with available data before actual operation, ensuring reliable control performance even when real-time data is limited. The control information is generated ahead of time based on historical patterns.
Solution Approach 2:
The patent uses copying by replicating and utilizing historical power consumption data from previous years that correspond to the same seasonal period. Instead of requiring extensive unique training data, the system copies and leverages patterns from historical periods (e.g., using January 2021 data to inform January 2022 control) to build a robust model with sufficient effective training samples.
2Speed
If real-time ESS control information is generated, then responsiveness to instantaneous changes is improved, but control time delays occur due to data insufficiency
Solution Approach 1:
The system performs preliminary generation of ESS control information by training the model in advance using historical data from the same seasonal period. This pre-training eliminates the need for time-consuming real-time training operations, allowing the system to respond quickly to instantaneous changes while avoiding control delays. The model is ready to operate immediately when deployed.
3Loss of time
If one year of power consumption data is used for training, then model preparation time is reduced, but the robustness of the AI model decreases
Solution Approach 1:
The patent applies preliminary action by utilizing power consumption data from multiple previous years (e.g., January 2020, January 2021, January 2022) to train the model in advance for the current period. This multi-year historical data approach ensures robust model training while maintaining efficient one-year deployment cycles, as the model is prepared beforehand using comprehensive historical patterns.
4Adaptability or versatility
If ESS control information is averaged over weekdays and holidays, then generalizability is improved, but the ability to capture specific day-type patterns is reduced
Solution Approach 1:
The patent applies segmentation by separating ESS control information into distinct weekday and holiday categories rather than averaging them together. This allows the system to maintain precise, tailored control strategies for each day type while still achieving generalizability through the reinforcement learning model that learns from both segments. Each segment retains its specific patterns.
Data Source
AI summary
A method of generating energy storage system (ESS) control information using reinforcement training results and a computing device for performing the method are provided. The method includes training a reinforcement training model for reducing peak load of an ESS using power consumption data corresponding to a first period in the past for a building to which the ESS is applied, generating ESS control information corresponding to the first period in the past by applying, to the trained reinforcement training model, the power consumption data corresponding to the first period in the past for the building, and converting the generated ESS control information corresponding to the first period in the past into ESS control information on a daily basis divided into weekdays and holidays and applying the ESS control information on a daily basis divided into weekdays and holidays as ESS control information corresponding to a second period at present.