Can You Trust Your Plant's Sensors? Machine Learning for Renewable Energy Assets

Energy management starts with data:  every optimisation, forecast or control decision in a solar or wind plant rests on what its sensors report. However, sensors can drift, fail and get miscalibrated and when they do, the damage can be missed.

‍

‍

This post summarises three pieces of research by our scientist Dr Fabio Rodríguez, developed during his PhD at the University of Seville with collaborators in Seville and Munster Technological University. Together they tell one story: how to detect bad data in renewable energy assets and how to plan the inspection once something is wrong.

‍

‍

The problem: digital twins learn from sensors

‍

‍

A digital twin is a virtual model of a physical plant. To stay accurate over the asset's life, it is periodically retrained on fresh sensor data. That is where it becomes important to avoid bad data. If a drifting sensor feeds the retraining, the twin quietly learns the wrong behaviour and every decision built on it gets worse.

‍

‍

The research proposes a simple principle: check every reading before it reaches the model. A fault detector flags suspicious values and a second model predicts what the reading should have been. The gap between the two (the offset) tells you how severe the problem is.

‍

‍

Case 1: a solar cooling plant

‍

The first paper, published in Computers in Industry (2023), applies the idea to a real solar cooling plant on the roof of the School of Engineering at the University of Seville. The plant uses a Fresnel solar field to heat water that powers an absorption chiller for air conditioning.

‍

The method combines two techniques:

‍

• A neurofuzzy detector classifies each sensor reading as normal, positively offset or negatively offset. Because it is built from readable fuzzy rules, engineers can inspect why it reached a decision.
• A recurrent neural network (LSTM) predicts the expected value of each sensor from its recent history, so the size of the error can be estimated.

‍

On test data, the detector reached about 97% accuracy for the solar irradiance sensor and about 90% for the flow sensor. In a full simulated working day, with errors in detection and prediction allowed to build up, detection accuracy stayed between 83% and 89% across sensors. The authors also report that the models need less than a second per reading, which matters for real-time use.

‍

‍

Case 2: wind turbines

‍

The second piece, presented at the SOCO 2022 conference (Springer), asks whether the same approach transfers to a different asset. The team used one year of data, at 10-minute intervals, from two real wind turbines, covering six sensors including wind speed, rotor speed, power and nacelle position.

‍

One LSTM network per sensor predicts the expected reading from recent history. A reading that departs too far from its prediction is flagged. Two findings stand out:

‍

• History matters. Giving the network more past readings consistently improved predictions, which shows that turbine behaviour is not random but depends on its recent state. With 30 past readings, the average prediction error stayed under 10% of each sensor's average value.
• Detection was strong. With simulated sensor faults, accuracy was above 80% on every sensor and close to perfect on several.

‍

The authors also mention  a limitation. When faults are detected early in the day, the system patches its history with predictions. Therefore,  small residual errors accumulate so accuracy drops later on. Their practical remedy is to physically check a sensor when faults repeat. 

‍

‍

Case 3: when something is wrong, who goes to look?

‍

Detecting a fault is half the job. Someone, or something, has to inspect. The third paper, published in Information Processing Letters (2025) with colleagues from the University of Texas at Dallas and the University of Seville, tackles the logistics: how should a battery-limited drone cover a set of segments, such as power lines or rows of solar tubes, flying out from and back to a base station?

‍

In general this kind of problem is NP-hard, meaning no efficient exact method is known. The paper shows that for segments lying on a line, it can be solved efficiently:

‍

• Fewest flights: a greedy strategy is provably optimal and runs in time proportional to the number of segments plus the number of tours.
• Shortest total distance: a dynamic-programming approach finds the optimal set of tours in polynomial time.
• ‍

This is theoretical work, with proofs rather than field trials. The authors point to horizontal segments laid out on a grid, as in the tube layout of solar plants, as a natural next step.

‍

‍

What this means for energy management

‍

Put the three together and you get a pipeline for the operational side of renewables:

‍

1. Trust the data: Detect faulty sensor readings and estimate their size before they contaminate models.
2. Reuse the method.: The same approach worked on a solar cooling plant and on wind turbines  and the authors designed it to adapt to other sensor-based systems.
3. Plan the response: Use efficient routing to send inspection drones where they are needed, within battery limits.

‍

‍

Read the papers

‍

• Rodríguez, F., Chicaiza, W. D., Sánchez, A., Escaño, J. M. Updating digital twins: Methodology for data accuracy quality control using machine learning techniques. Computers in Industry 151 (2023). https://doi.org/10.1016/j.compind.2023.103958
• Rodríguez, F., Chicaiza, W. D., Sánchez, A. J., Escaño, J. M. Neural Networks Techniques for Fault Detection and Offset Prediction on Wind Turbines Sensors. In: SOCO 2022, Lecture Notes in Networks and Systems 531, pp. 534-543 (2023). https://doi.org/10.1007/978-3-031-18050-7_52
• Bereg, S., Díaz-Báñez, J.-M., Kasiuk, A., Pérez-Cutiño, M.-A., Rodríguez, F. Covering segments on a line with drones. Information Processing Letters 188 (2025). https://doi.org/10.1016/j.ipl.2024.106540

‍

The wind turbine and solar plant data sets are publicly available on Dr Fabio Rodriguez's GitHub (github.com/fabio-rodriguez).

‍

The fault detection work received funding from the European Union's Horizon 2020 research and innovation programme under grant agreement no. 958339 (DENiM project).

‍