AI methods for heat and power generation with solar thermal collector systems


During the operation of concentrating solar power (CSP) systems, a large amount of measurement data accumulates that has so far only been used for simple control purposes. Secondary information that allows conclusions about the status of system components, the quality of the control system and the early detection of malfunctions has remained unused until now. The project investigates and tests AI-based methods for monitoring and optimizing solar thermal plants.

Project description

Due to the complex interaction in concentrating solar thermal systems between the extensive optical systems and the thermal-hydraulic high-temperature processes, the degree of automation is still comparatively low. However, the automation and optimization of system operation holds significant potential for further cost reduction.

Within this project, existing ground-based and airborne optical remote sensing systems for monitoring environmental conditions and the condition of power plants will be optimized using AI methods. Innovative semi-supervised and self-supervised methods are used, which can work with extensive unlabeled or only partially labelled data. The measurement results of the remote sensing systems are combined with the thousands of sensor channels of the plants (temperature, mass flow, collector position, pressure, etc.). Due to the amount of data and the variety of overlapping physical effects, an optimal utilization of the existing data is not possible with classical methods. For this purpose, the vision of an intelligent "O&M assistant" is being pursued, which continuously accesses and learns from all available information. Integrated into the control system, such an O&M assistant can make a significant contribution to the complete automation of CSP systems and thus increase system efficiency and reliability. Germany's positioning in this innovative condition monitoring and automation field also opens up a wide range of global marketing opportunities that promote the shift to renewable energies.

Research contribution

When evaluating the solar thermal operating data, variables must be taken into account that drastically complicate the application of classic machine learning methods. For example, the measurement data are subject to weather and seasonal fluctuations and can show peculiarities that are specific to each plant and geographical position. To nevertheless use the acquired sensor signals for tasks such as fault detection or the estimation of pollution levels, models are required that are largely insensitive to such variabilities. The development of models with invariance with respect to data distribution, data syntax and the desired task is a central research interest. An important aspect is the extraction of descriptors and development of distance measures for temporally and spatio-temporally resolved data.

The developed methods and approaches will be evaluated and analyzed with regard to their suitability. This requires, among other things, the use of suitable simulation environments, which are also in the research interest of fortiss. In connection with CO simulation environments, the connection and integration of real components is also essential in order to draw conclusions about the actual system behavior. Possibilities of injecting potential faults into the existing system are also being investigated here.

Project goals fortiss is researching:

  1. Optimized AI algorithms for operational data analysis
    Spatio-temporally correlated operational data is an application-specific challenge of the project. Existing AI methods will be adapted or combined to the extent that methods are usable and tested for application in solar problems. Proof of functionality in WP 5 demonstrates the added value of the new methods.
  2. Anomaly detection in solar power plants using AI
    The detection of degradations or drifts in the abundance of operational data is complicated in solar power plants by the constantly changing operating points (solar irradiance, ambient temperature,...). The identification of creeping effects can only be detected by the human plant operator when they have grown to a major failure. Therefore, new approaches based on AI are being developed to analyze plant operating data to identify anomalies, patterns and trends relevant to plant operation and lifetime. The operational measurement data transmitted by the plant will be augmented with additional generated "smart data" and the use of the self-learning plant models. It is envisioned that the algorithm will be able to classify collectors and recognize temporal as well as spatial patterns in the solar field.
  3. AI-based operational assistance systems for decision-making processes
    Algorithm to support the power plant in daily and longer-term decisions by means of an operation assistance system based on AI data evaluation, which allows interaction with the user, visualizes the processed data and derives power plant-specific recommendations from the AI results.
    At a minimum, it is envisioned that the algorithm will be able to make meaningful longer-term recommendations (for example, receiver replacement) and medium-term recommendations (for example, tracking recalibration or hydraulic readjustment).
  4. AI-based plant control for process steam plants
    Direct expansion processes can be run at optimal operating points depending on ambient conditions. Due to limited metrological insight, the current plant condition is not directly measurable at reasonable cost, but is critical for optimization. The goal is to build an AI-based process monitoring system that automatically determines the system state, taking into account specific sensors and histories. This information is then to be incorporated into an AI-supported operation control logic, which ensures that the system is operated optimally at all times without human intervention. The operation control logic will also draw on radiation prediction system. The operation control logic will be able to make decisions based on the predictive values of the damaged or missing sensors, thus avoiding any downtime that may be caused by the dependence of the control algorithm on some sensor data.

Project duration

04.07.2022 - 03.07.2025

Dr. Markus Duchon

Your contact

Dr. Markus Duchon

+49 89 3603522 30

Project partner