Energy Forecasting Model Transfer for Cold-Start Buildings
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Smart Energy Management Systems face challenges in accurately forecasting energy consumption in newly instrumented buildings due to reliance on historical data, leading to inaccurate energy estimations and performance gaps.
Innovation Solution
The method employs Few-Shot Learning through a Cold Start approach, pretraining a model with generic data and fine-tuning it using limited data from similar buildings, leveraging time series analysis and recurrent neural networks to predict energy consumption or production, while freezing most layers and using regularization techniques for improved adaptability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If historical data from newly instrumented buildings is used for training energy forecasting models, then the models can be trained with available data, but the forecasts are inaccurate due to limited and low-quality data
Solution Approach 1:
The patent applies preliminary action by pre-training the neural network model on historical energy data from existing buildings before deploying it to newly instrumented buildings. This allows the model to learn from abundant historical data in advance, so that when deployed to new buildings with limited data, it already possesses predictive capabilities without requiring extensive local training data
Solution Approach 2:
The patent uses copying by transferring the trained model weights and parameters from source buildings (with abundant historical data) to target buildings (with limited data). The model learned from source building data is copied and adapted to predict energy consumption in target buildings, eliminating the need to collect and train on large amounts of target building data
2Adaptability or versatility
If models are trained from scratch for each new building, then the models can be customized to specific building characteristics, but the training requires large amounts of high-quality historical data that are unavailable in early deployment phases
Solution Approach 1:
The system performs preliminary training on diverse building data from multiple source buildings before deployment. This pre-training phase allows the model to learn general energy consumption patterns across different building types, and then this pre-trained model is transferred to target buildings where it can be fine-tuned with minimal local data, achieving adaptability without requiring large amounts of building-specific training data
Solution Approach 2:
The patent implements universality by creating a single model architecture that can be applied across multiple different building types and contexts. The model is trained on diverse data from various source buildings, making it universally applicable to different target buildings. This universal model can be adapted to specific building characteristics through transfer learning without requiring separate models for each building
3Productivity
If classical machine learning or deep learning approaches are used that rely on historical data, then the models can process available data, but recently instrumented buildings without sufficient historical data cannot achieve their performance goals
Solution Approach 1:
The system performs model training and validation in advance using historical data from existing buildings before the model is deployed to newly instrumented buildings. This preliminary training action ensures that when the model is deployed, it immediately achieves high performance without requiring a lengthy data collection and training period at the target building, thus eliminating the performance gap during early deployment phases
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Method for training at least one model able to predict a power consumption or production of at least one electric equipment, also called target, said method comprising: (a) obtaining (S1) time series data representing the evolution of the power consumption or production of said target over a first period of time, (b) comparing (S2) said target time series data to known time series data representing the evolution, over a second period of time, of the power consumption or production of known electric equipments, said second period of time being greater than the first period of time, to determine the k known time series data that are the most similar to the target time series data, (c) training (S3) of a first prediction model or backbone for each of said k known electric equipments, said backbone being able to predict the evolution over time of the consumption or the production of the corresponding known electric equipment, said backbone being trained on the corresponding time series data over the second period of time, (d) training (S4) at least one second prediction model, called target model, by fine tuning at least one of the first trained prediction model on the time series data of said target over the first period of time.