Air conditioner control based on prediction from classification model
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The existing air conditioning control systems face challenges in maintaining prediction accuracy due to the time lag in updating the learning model, which results in inadequate responsiveness to sudden temperature changes, leading to deteriorated prediction performance.
Innovation Solution
The system allows edge terminals to locally update the learning model using prediction results and request relearning from the cloud server when cumulative prediction errors exceed a threshold, reducing the frequency of communications and enhancing responsiveness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the learning model is updated frequently by the cloud server, then the prediction accuracy is improved, but the communication frequency and system complexity increase
Solution Approach 1:
The system divides the model update function into two parts: the cloud server performs initial model training and periodic updates, while edge terminals perform local adaptive updates using sequential machine learning. This segmentation allows frequent local adaptations without requiring constant cloud communication, thus maintaining prediction accuracy while reducing system complexity.
Solution Approach 2:
The cloud server performs preliminary model training in advance and distributes the trained model to edge terminals. The edge terminals then use this pre-trained model as a baseline and perform incremental local updates. This preliminary action reduces the need for frequent cloud communication while maintaining accurate predictions through local adaptations.
2Device complexity
If the learning model is updated by the cloud server at preset intervals, then the system complexity is reduced, but the responsiveness to sudden temperature changes deteriorates
Solution Approach 1:
The system transitions from static periodic updates to dynamic adaptive updates. Edge terminals continuously monitor prediction errors and trigger model relearning when errors exceed a threshold, allowing the system to adapt dynamically to sudden temperature changes while maintaining relatively simple architecture.
Solution Approach 2:
The system implements feedback mechanisms where edge terminals monitor prediction accuracy and send requests to the cloud server when performance degrades. This feedback loop enables responsive adaptation to environmental changes while keeping the cloud server involved only when necessary, balancing simplicity and adaptability.
3Adaptability or versatility
If edge terminals perform local sequential machine learning, then the responsiveness to temperature changes is improved, but the computation load at the edge increases
Solution Approach 1:
Instead of performing full model retraining at the edge, the system implements sequential machine learning that performs partial updates incrementally. This approach provides the responsiveness benefits of local learning while significantly reducing the computational burden compared to complete model retraining, making edge computation feasible.
4Reliability
If the cloud server collects log information from multiple users, then the model generalization is improved, but the communication data volume increases
Solution Approach 1:
The system extracts only essential features and aggregated statistics from user logs rather than transmitting complete raw data to the cloud. This extraction approach maintains model generalization by preserving key patterns while significantly reducing communication data volume between edge terminals and the cloud server.
Data Source
AI summary
A prediction method implemented by a computer, the method includes: receiving a classification model from a server, the classification model being a model for classifying logs of an electronic device into two or more classes, the server being a computer configured to distribute the classification model; calculating, with respect to different time points, a prediction error by using a predicted value outputted by the classification model and an actual measured value observed at each of the different time points; performing sequential machine learning for the classification model to have the prediction error satisfy a certain condition; and when a cumulative sum with respect to the prediction error of the sequential machine learning is equal to or greater than a threshold, requesting the server apparatus to relearn the classification model.