Power Prediction Model Training with Multi-Time-Scale Data Fusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current power prediction solutions face challenges in accurately predicting solar power generation and load consumption due to data loss from random sampling and the need for manual hyperparameter optimization, as well as the impact of domain knowledge constraints on model accuracy.
Innovation Solution
A model training method that utilizes historical power and meteorological data from multiple sources and time scales, employing different models for global and local representations, with adaptive weights to optimize prediction precision.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of time
If random sampling is performed in frequency domain to reduce calculation complexity, then computation time is reduced, but information loss occurs in the input data
Solution Approach 1:
The patent segments the input data processing into multiple parallel branches: one branch processes sampled frequency domain data for fast computation, while another branch processes complete multi-time scale data for comprehensive information extraction. This segmentation allows the system to balance speed and accuracy by combining results from different processing paths.
Solution Approach 2:
The patent introduces multi-time scale analysis as an additional dimension to the traditional single-time-scale frequency domain sampling. By analyzing data at multiple time scales (different sampling rates and windows), the system recovers information that would be lost in conventional single-scale sampling while maintaining computational efficiency through selective processing.
2Measurement precision
If manual intervention is used to optimize hyperparameters for balancing operation performance and model precision, then prediction accuracy is improved, but operation complexity increases
Solution Approach 1:
The patent implements self-service through automated hyperparameter optimization using grid search and cross-validation. The system automatically searches for optimal hyperparameter combinations and validates them, eliminating the need for manual intervention while maintaining high prediction accuracy. This is achieved through programmatic exploration of parameter spaces and automated model evaluation.
3Reliability
If domain knowledge is added as loss constraints to restrict and optimize model learning, then model reliability is improved, but accuracy is affected by precision of bound fitting functions
Solution Approach 1:
The patent employs dynamic loss weighting that adapts the influence of domain knowledge constraints during training. Instead of using fixed loss constraints, the system dynamically adjusts the weight of physical constraint losses based on training progress and model performance, allowing the model to learn from domain knowledge without being overly constrained by potentially imprecise bound fitting functions.
Solution Approach 2:
The patent changes the parameter representation of domain knowledge constraints from fixed boundary values to flexible, learnable parameters. By representing physical constraints as learnable parameters that can be optimized during training, the system maintains reliability through domain knowledge while improving accuracy through adaptive parameter adjustment rather than relying on pre-defined fitting function precision.
Data Source
AI summary
This application discloses a model training method, a power prediction method, and a device. A model includes at least a first model and a second model. The method includes: obtaining a dataset including historical power data and historical meteorological data in preset duration; determining a weight of a loss function of each model based on the dataset; constructing a target loss function, where the first model and the second model may be used to represent input data, and data represented by the first model is not completely the same as data represented by the second model; and finally training the models based on the target loss function. More complete information is obtained from different modal data and different time scales based on a currently available data source, to optimize the learning model and achieve higher prediction precision.