EP-4741954-A1 - COMPUTER-IMPLEMENTED METHOD FOR DETERMINING A COMMON TRAINED MACHINE LEARNING MODEL FOR A FLEET OF POWER PLANTS

EP4741954A1EP 4741954 A1EP4741954 A1EP 4741954A1EP-4741954-A1

Abstract

The invention relates to a computer-implemented method for determining a common trained machine learning model for a fleet of power plants, comprising the steps: a. Providing a plurality of training data sets for a plurality of respective power plants (S1); wherein the plurality of training data sets is combined in a heterogeneous training data set; b. Providing a machine learning model (S2); c. Training the machine learning model on the heterogeneous training data set (S3); and d. Providing the common trained machine learning model (S4). Further, the invention relates to a corresponding computer program product and technical unit.

Inventors

BOGDOLL, DIETER
Depeweg, Stefan
Schöner, Holger
SUN, YIZE
Swazinna, Phillip
TIETZ, CHRISTOPH
Tokic, Michel
von Beuningen, Anja

Assignees

Siemens Aktiengesellschaft

Dates

Publication Date: 20260513
Application Date: 20241111

Claims (10)

Computer-implemented method for determining a common trained machine learning model for a fleet of power plants, comprising the steps: a. Providing a plurality of training data sets for a plurality of respective power plants (S1); wherein the plurality of training data sets is combined as a heterogeneous training data set; b. Providing a machine learning model (S2); c. Training the machine learning model on the heterogeneous training data set (S3); and d. Providing the common trained machine learning model (S4).
Computer-implemented method according to claim 1, wherein at least one power plant of the plurality of power plants is a gas turbine, wherein preferably each power plant of the plurality of gas turbines is a gas turbine.
Computer-implemented method according to claim 1 or claim 2, wherein each training data set of the plurality of training data sets comprises sensor data, wherein the sensor data is determined by means of at least one sensor.
Computer-implemented method according to claim 3, wherein a respective identifier is assigned to each sensor of the at least one sensor and a respective value of the at least one sensor is encoded as a token.
Computer-implemented method according to any of the preceding claims, wherein the machine learning model is a neural network, preferably a transformer or a diffusion model.
Computer-implemented method according to any of the preceding claims, further comprising - Fine-tuning the common trained machine learning model using a fine-tuning approach based on an additional training data set from an additional power plant, wherein the fine-tuning approach utilizes preferably a Low-Rank Adaptation; and/or - Prompt-tuning the common trained machine learning model.
Computer-implemented method according to any of the preceding claims, further comprising - Applying the common trained machine learning model, the fine-tuned common machine learning model or prompt-tuned common machine learning model on an input data set of an input power plant.
Computer-implemented method according to any of the preceding claims, further comprising - Outputting the common trained machine learning model, the fine-tuned common machine learning model, any other input and/or output data; - Storing the common trained machine learning model, the fine-tuned common machine learning model, any other input and/or output data; - Displaying the common trained machine learning model, the fine-tuned common machine learning model, any other input and/or output data; - Transmitting the common trained machine learning model, the fine-tuned common machine learning model, any other input and/or output data; and/or - Evaluating the common trained machine learning model, the fine-tuned common machine learning model, any other input and/or output data.
A computer program product directly loadable into an internal memory of a computer, comprising software code portions for performing the steps according to any one of the preceding claims when said computer program product is running on a computer.
Technical unit for performing the method according to any one of claims 1 - 8.

Description

1. Technical field The present invention relates to a computer-implemented method for determining a common trained machine learning model for a fleet of power plants. Further, the invention relates to a corresponding computer program product and technical unit. 2. Prior art Machine learning is well known from the prior art in context of Artificial Intelligence. According to which, the Artificial Intelligence systems are software programs whose behavior is learned from data instead of being explicitly programmed. The learning process is called "training" which requires plenty of training data and significant computational resources. Thereby, the trained machine learning model solves a specific task for which it was trained, such as data classification. In other words, regarding classification, the trained machine learning models are configured to classify the data elements. The trained Machine-Learning-Models are increasingly used and gain in importance in diverse technical fields. For example, they can be used for object detection in autonomous vehicles. Moreover, power plants (e.g. gas turbines) or industrial plants can be nowadays controlled by means of the trained Machine-Learning-Models. This control allows for significantly reduced emissions of the power plants or increased performance of the industrial plants. Known control mechanisms comprise the setting of e.g. the fuel fractions during operation of the power plants in order to optimize the trade-off between NOx emission and acoustic dynamics. Usually, the Machine-Learning-Models or Artificial Intelligence systems are not learned or optimized on their target technical systems, such as the aforementioned autonomous vehicles, industrial plants or power plants. To the contrary, they are trained on a separate computing unit or simulator due to safety concerns. The Machine-Learning-Model can be configured to model the effect of the control variables (e.g. the fuel fractions) on the emissions of the power plants. After the control strategy is learned on the sperate computing unit or simulator it is then deployed on the target technical system to e.g. reduce the emissions. According to prior art approaches, usually, the Machine-Learning-Models are trained for each technical system, such as power plants, separately. In other words, one Machine-Learning-Model (or ensemble of those) is trained for each instance within a fleet of similar power plants. For instance, data recording of a particular power plant is stored and then based on this data a Machine-Learning-Model is trained, modelling for example state-transitions, emissions (NOx, CO) and/or dynamic frequency bands (IFD, HFD, LFD). Subsequently, a control strategy is derived utilizing the trained Machine-Learning-Model, such as minimizing emissions and/or frequency dynamics. The disadvantage is that the prior art approaches are time-consuming and cost intensive. It is therefore an objective of the invention to provide a computer-implemented method for determining a common trained machine learning model for a fleet of power plants, which is more efficient and reliable. 3. Summary of the invention This problem is according to one aspect of the invention solved by a computer-implemented method for determining a common trained machine learning model for a fleet of power plants, comprising the steps: a. Providing a plurality of training data sets for a plurality of respective power plants; wherein the plurality of training data sets is combined as a heterogeneous training data set;b. Providing a machine learning model;c. Training the machine learning model on the heterogeneous training data set; andd. Providing the common trained machine learning model. Accordingly, the invention is directed to a computer-implemented method for determining a common trained machine learning model for a fleet of power plants. The machine learning model can be equally referred to as model based on or utilizing machine learning or Machine-Learning-Model. The common trained machine learning model can be understood as one single or one joint machine learning model. This machine learning model is specified, used for training and the generation of the trained machine learning model for the plurality of power plants or whole set of power plants. Preferably, the fleet of power plants is provided in the form of a plurality of different or diverse gas turbines. The trained common machine learning model can be designed as a simulation model which models for example state-transitions, emissions (e.g. NOx and CO) and/or dynamic frequency bands (e.g. IFD, HFD and LFD). After the control strategy is learned on the simulator it can then be deployed on the target power plant. Alternatively, the machine learning model can be configured for anomaly detection. Therefore, in the first step training data sets are provided for the power plants in the form of a heterogeneous training data set. In the second step, the machine learning model is provided