Demand Charge Forecasting with LSTM and RL for Battery Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional behind-the-meter (BTM) energy management systems face challenges in accurately forecasting demand charge thresholds, leading to inefficient demand peak shaving and increased electricity bills for commercial and industrial customers due to significant forecasting errors.
Innovation Solution
A computer-implemented method and system that generates a forecast model using historical demand charge threshold (DCT) profiles to predict the next month's DCT value, employing machine learning tools like K-mean clustering and Ridge regression, which then controls a battery storage system to prevent utility power from exceeding the forecasted DCT, thereby optimizing charge and discharge schedules.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional load forecasting methods are used to predict demand charge thresholds, then the forecasting process is simple to implement, but the forecasting accuracy is poor leading to significant errors
Solution Approach 1:
The forecasting approach is segmented into two distinct components: a supervised learning model that captures temporal dependencies using historical sequences, and a reinforcement learning agent that optimizes demand charge thresholds based on learned patterns. This segmentation allows each component to specialize in specific aspects of forecasting, improving overall accuracy while maintaining manageable complexity through modular architecture.
Solution Approach 2:
The system employs a composite forecasting model that integrates supervised learning (LSTM networks) and reinforcement learning (deep Q-networks) components. This composite approach combines the strengths of both methodologies: supervised learning provides accurate baseline predictions from historical data, while reinforcement learning optimizes thresholds by learning from environmental feedback, achieving superior forecasting accuracy compared to either method alone.
2Loss of energy
If a battery storage system is used to shave demand peaks, then the demand charge cost is reduced, but the system complexity and control difficulty increase
Solution Approach 1:
The reinforcement learning agent operates autonomously to manage battery storage systems, making discharge decisions based on learned patterns from historical data and real-time environmental feedback. The system self-optimizes demand charge threshold selection and battery dispatch strategies without requiring manual intervention or complex centralized control, reducing operational complexity while maximizing demand charge reduction through adaptive, autonomous decision-making.
3Power
If optimal charging and discharging schedules are implemented, then the maximum demand peak is minimized, but the requirement for accurate load forecasting increases system difficulty
Solution Approach 1:
The supervised learning component performs preliminary action by pre-processing historical load data to extract temporal patterns and dependencies before the reinforcement learning agent makes optimization decisions. This preliminary analysis of historical sequences enables the RL agent to focus on threshold optimization rather than raw pattern recognition, reducing the forecasting difficulty while maintaining the ability to minimize demand peaks through informed, data-driven scheduling decisions.
Data Source
AI summary
A computer-implemented method, system, and computer program product are provided for demand charge management. The method includes receiving an active power demand for a facility, a current load demand charge threshold (DCT) profile for the facility, and a plurality of previously observed load DCT profiles. The method also includes generating a forecast model from a data set of DCT values based on the current load DCT profile for the facility and the plurality of previously observed load DCT profiles. The method additionally includes forecasting a monthly DCT value for the facility using the forecast model. The method further includes preventing actual power used from a utility from exceeding the next month DCT value by discharging a battery storage system into a behind the meter power infrastructure for the facility.