Control of Technical Systems Using Trainable Time-Series Filtering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current control methods for complex technical systems, such as wind turbines and power grids, face inefficiencies in predicting system behavior due to the need for extensive preprocessing of time series data and decreased training success with longer sequences, especially when using recurrent neural networks.
Innovation Solution
A method involving a trainable digital filter and a machine learning routine, where the digital filter extracts relevant operating parameter features and the machine learning routine predicts target operating parameters, allowing for efficient and accurate predictions by training both components together, particularly using convolutional neural networks that scale well with longer sequences.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If recurrent neural networks are used to evaluate time series data, then prediction capability is maintained, but training success decreases with longer time series
Solution Approach 1:
The patent segments the time series evaluation process into two distinct components: a digital filter that performs preprocessing to extract relevant features and reduce data dimensionality, and a machine learning routine that performs prediction based on the filtered data. This segmentation allows each component to specialize, with the filter handling the computationally intensive preprocessing and the ML routine focusing on prediction, thereby improving training efficiency especially for long time series while maintaining prediction accuracy
2Productivity
If time series data are preprocessed to extract relevant patterns, then data volume is reduced and prediction efficiency is improved, but preprocessing requires expert design and is time-consuming
Solution Approach 1:
The digital filter in the patent is designed to be trainable, allowing it to automatically learn and adapt to the specific characteristics of the time series data through training. This self-learning capability eliminates the need for manual expert-designed preprocessing, as the filter automatically identifies and extracts relevant features from the data, significantly reducing the time and expertise required for preprocessing while improving prediction efficiency
3Measurement precision
If multiple operating parameter values are evaluated for prediction, then prediction accuracy is improved, but computational complexity increases
Solution Approach 1:
The digital filter extracts only the most relevant features and patterns from the multiple operating parameter values, separating the essential predictive information from the redundant or less important data. This extraction process reduces the dimensionality of the input data while preserving the critical information needed for accurate prediction, thereby reducing computational complexity without sacrificing prediction accuracy
Data Source
Figure 1
Figure 2
AI summary
In order to control a technical system (TS), e.g. of a wind turbine, a temporal sequence of operating parameter values (BP) of the technical system (TS) is continuously recorded and continuously converted into a sequence of filtered signal values (GS) by a trainable digital filter (DF). The sequence of the filtered signal values (BS) is supplied to a mechanical learning routine (NN) which derives prediction values (PZ) therefrom for a target operating parameter (ZBP). The digital filter (DF) and the mechanical learning routine (NN) are trained to reduce a distance (D) between derived prediction values (PZ) and temporally corresponding, actually recorded values of the target operation parameter (ZBP). The prediction values (PZ) for controlling the technical system (TS) are then emitted.