Data-Driven Controller Configuration for Building Energy and Comfort
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Building management systems face challenges in optimizing energy consumption and comfort due to the complexity of parameters involved, with existing model predictive control approaches requiring significant modeling effort and not considering all relevant information.
Innovation Solution
A computer-implemented method using a two-step data-driven approach with a first model predicting future target variables and a second model trained via offline reinforcement learning to configure a controller, balancing energy consumption and comfort through a reward-based strategy, incorporating past and future sensor data and external forecasts.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If model predictive control approaches use simplified building models to predict control variables, then future room conditions can be predicted and optimization over control settings is enabled, but the models do not consider all relevant information and require significant modeling effort and calibration
Solution Approach 1:
The patent replaces traditional physics-based building models with machine learning models that learn building behavior patterns directly from operational data. This substitution eliminates the need for complex manual modeling and calibration while capturing all relevant information present in the operational data, thereby resolving the contradiction between control optimization performance and modeling effort.
Solution Approach 2:
The machine learning models automatically learn and adapt to the specific building's characteristics through self-training on operational data. This self-service approach eliminates the need for external experts to perform manual model calibration and ensures the models consider all relevant information present in the building's operational history, resolving the contradiction between optimization performance and modeling complexity.
2Ease of manufacture
If traditional MPC approaches use simplified models with limited information, then implementation is easier, but the models fail to consider all relevant information for optimal control
Solution Approach 1:
The patent creates a universal data-driven framework that can process and integrate multiple types of building information (occupancy patterns, weather data, energy consumption, sensor readings) through a single machine learning model architecture. This multi-functional approach maintains implementation ease while eliminating information loss by universally handling all relevant data types.
Solution Approach 2:
The patent transforms the modeling approach from using simplified physical parameters to using data-driven parameters learned from operational data. This parameter transformation allows the model to consider all relevant information while maintaining implementation ease through standardized machine learning pipelines and automated training procedures.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The invention refers to a computer-implemented method for configuring a controller (CO) for a technical system, where the controller (CO) controls the technical system based on an output data set (OS) determined by the controller (CO) for an input data set (IS), wherein the method comprises the following steps: - training a first data driven model (SM) with training data comprising several pre-known input data sets (IS) and corresponding pre-known output data sets (OS) for the respective pre-known input data sets (IS), where the first data driven model (SM) predicts respective future values of one or more target variables (tv) for one or more subsequent time points (tp) after a current time point; - training a second data driven model (PO) with said training data using reinforcement learning with a reward (RW) depending on the respective future values of said one or more target variables (tv) which are predicted by the trained first data driven model (SM), where the trained second data driven model (PO) is configured to determine the output data set (OS) for the input data set (IS) within the controller (CO).