Informer-Based Load Decomposition for Accurate Complex Load Disaggregation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing non-intrusive load decomposition methods face challenges such as low generalization ability, data imbalance, and low comprehensiveness in evaluation indexes, leading to inaccurate decomposition of complex loads, particularly with deep learning-based approaches.
Innovation Solution
A non-intrusive load decomposition method utilizing an Informer model coding structure, which includes feature extraction, processing, and mapping parts, employing one-dimensional convolutional layers, self-attention mechanisms, and positional embedding to improve feature representation and reduce complexity, while using a specific loss function and masking technique for enhanced performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional deep learning algorithms (automatic encoders, convolution neural networks, cyclic neural networks) are used for non-intrusive load decomposition, then the decomposition accuracy can be improved, but the model generalization ability deteriorates and training data imbalance problems arise
Solution Approach 1:
The Informer model is designed as a universal architecture that can handle multiple types of load decomposition tasks simultaneously. By incorporating flexible input processing and a unified attention mechanism, the model adapts to different load types and data conditions without requiring separate specialized models, thereby improving generalization ability while maintaining decomposition accuracy
Solution Approach 2:
The model dynamically adjusts parameters such as attention weights, sequence lengths, and feature extraction parameters based on the specific characteristics of the input data. This parameter adaptability allows the model to maintain high decomposition accuracy across different scenarios while avoiding overfitting to specific training data distributions
2Measurement precision
If complex deep learning models are used to achieve high decomposition accuracy, then the decomposition precision improves, but the computational complexity and training requirements increase
Solution Approach 1:
The Informer model segments the processing of long sequences into manageable blocks through its encoder-decoder architecture with attention mechanisms. This segmentation allows the model to process complex load decomposition tasks in distributed steps, reducing the immediate computational burden while maintaining overall accuracy through coordinated processing across segments
Solution Approach 2:
The model introduces intermediate feature representations and attention mechanisms that act as mediators between the raw input data and final decomposition results. These intermediaries simplify the computational path by extracting and prioritizing key features, thereby reducing the complexity required in subsequent processing stages while preserving decomposition accuracy
3Adaptability or versatility
If long sequence data is processed to improve decomposition performance, then the generalization ability improves, but the memory and time complexity increase
Solution Approach 1:
The Informer model employs periodic attention mechanisms that selectively focus on relevant time steps and features at regular intervals rather than processing every element in the sequence. This periodic action enables the model to capture long-term dependencies for improved generalization while significantly reducing the computational time and memory requirements compared to processing entire long sequences uniformly
Data Source
AI summary
The present disclosure relates to non-intrusive load decomposition techniques, and aims at providing a non-intrusive load decomposition method based on an Informer model coding structure. The method includes: preprocessing power data in electricity loads, and forming sample pairs according to a time sequence of total power and a time power sequence of a single electricity load; building a training model with reference to an Informer model, the training model including a feature extraction part, a feature processing part, and a feature mapping part which are arranged in sequence; initializing parameters of the training model, and selecting an appropriate activation function and loss function; training the training model using preprocessed sample data; and inputting a total power curve to the trained model, and conducting decomposition to obtain a power curve of a single load.