Integrated Energy Dispatch Using LSTM Multi-Energy Flow Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for multi-energy flow calculation in integrated energy systems face challenges such as high nonlinearity, complex coupling relationships, and low calculation accuracy, especially when dealing with large-scale systems.
Innovation Solution
A method and system that utilize a double-hidden-layer LSTM neural network for nonlinear regression of electrical power flow and a single-hidden-layer LSTM neural network for hydraulic models, combined with error compensation and a mechanism-driven linear thermal model, to improve calculation accuracy and efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional independent modeling method and Newton-Raphson solution algorithm are used for multi-energy flow calculation, then the calculation can be performed based on established models, but the solution time is long and convergence is difficult for large-scale IES
Solution Approach 1:
The patent segments the complex multi-energy flow calculation into separate electrical power flow model and thermal-hydraulic model components. Each model is trained independently using LSTM neural networks, allowing parallel processing and reducing overall solution time while maintaining calculation accuracy through dedicated specialized models for each energy domain
Solution Approach 2:
The patent replaces the traditional Newton-Raphson iterative numerical method with LSTM neural network-based direct calculation. The LSTM models are trained offline to learn the nonlinear relationships in energy flow, enabling fast online prediction without iterative convergence issues, thus eliminating the time-consuming iteration process while preserving accuracy
2Loss of time
If linearization or mathematical statistics methods are used for mechanism-driven modeling, then the calculation time is shortened, but the calculation accuracy is low and system operation analysis is affected
Solution Approach 1:
The patent substitutes linearization and mathematical statistics methods with LSTM neural network-based nonlinear regression. The LSTM architecture captures complex nonlinear relationships in energy flow without requiring system simplification or linearization, achieving both fast calculation and high accuracy through learned patterns from training data
Solution Approach 2:
The patent changes the modeling approach from parameter-based linearization to data-driven nonlinear parameter learning. The LSTM networks learn optimal parameters and relationships directly from training data, adapting to the actual system behavior without being constrained by linearization assumptions, thereby maintaining accuracy while ensuring fast computation
3Measurement precision
If AC power flow calculation method is used for electrical power system, then the calculation is comprehensive, but the high nonlinearity makes the solution process complex and convergence speed slow
Solution Approach 1:
The patent replaces the complex iterative AC power flow calculation with an LSTM neural network model trained to predict power flow results directly. The LSTM captures the nonlinear relationships learned from AC power flow training data, providing comprehensive accurate results through direct prediction without the complex iterative solution process
Solution Approach 2:
The patent performs preliminary training of the LSTM model using comprehensive AC power flow calculation data before actual operation. This offline training phase pre-computes the complex relationships, so that during online operation, the model provides fast predictions without repeating the complex iterative calculation process
4Measurement precision
If traditional modeling methods are used for multi-energy flow calculation, then the model can be established based on system structure, but with the increase of IES scale, the computational complexity is gradually increased
Solution Approach 1:
The patent segments the large-scale IES into modular electrical and thermal-hydraulic components, each with its own LSTM model. This modular approach allows the system to scale by adding more modular units without proportionally increasing overall computational complexity, as each module is processed independently through its specialized model
Solution Approach 2:
The patent changes from traditional explicit modeling that scales linearly with system size to LSTM-based implicit modeling where computation time remains relatively constant. The neural networks learn the system characteristics during training, enabling fast inference that does not increase proportionally with system scale
Data Source
AI summary
A method and system for performing analysis and calculation of multi-energy flow for integrated energy system (IES), comprising establishing electrical power flow model, hydraulic model and thermal model, to form preliminary model of IES; constructing DHL-LSTM neural network for nonlinear regression of electrical power flow model, construct SHL-LSTM neural network for nonlinear regression of hydraulic model, and finding optimal parameters of two neural networks; training the two models, and adding error compensations into the two models; simplifying the thermal model, to obtain mechanism-driven linear thermal mode; and embedding mechanism-driven linear thermal model into error-compensated hydraulic model and error-compensated electrical power flow model respectively, to form and use final model of the IES to perform analysis and calculation of multi-energy flow. The present invention fully considers the coupling property inside the system, and avoids the situation that the convergence speed is slowed down when solving high nonlinear problems.