Part of the parabolic trough section of the NOOR I CSP project in Morocco ©MASEN
A groundbreaking new AI simulation tool that dramatically improves power plant operations for parabolic Trough-type Concentrated Solar Power (CSP) plants has been created by the University of Wisconsin-Madison’s Michael Wagner, who helped develop the widely used System Advisor Model (SAM), and former graduate student and AI expert Matthew Tuman.
Their study, Neural-network-driven dynamic simulation of parabolic trough solar fields for improved CSP plant operation, published in the journal Solar Energy, describes this neural network-based model for managing Trough CSP plants, accounts for much more granular detail, while reducing simulation time by orders of magnitude, compared to traditional methods.
“This neural network, instead of taking about 5 seconds per time step to solve the full field model, it takes 0.05 seconds to evaluate,” said Wagner in a call from the US.
“And you have to solve this for thousands and thousands of individual evacuated tube receivers spread throughout the huge solar field in a Trough plant. So we see these tremendous speed-ups. It also allows us to capture the same level of detail you need to replicate realistic plant behaviors, but you can do it in sub-second simulations.”
While previous simulations have focused primarily on solar field performance metrics, this new approach dives deeper by modeling all the thousands of mechanical components, including control valves, pumps, piping, and the heat transfer fluid (HTF) expansion system.
South Africa’s Bokpoort-CSP-plant-with-parabolic-trough-collectors-©SENER
“We’re trying to capture a lot more detail than has been captured in prior models,” he said.
“When you have a really complex plant, like in Trough CSP, there’s so much going on, there are so many different sub-systems, and they’re all talking to each other by sharing common heat transfer fluid, or flows across boundaries of these subsystems, you need complex models to capture all those dynamics. AI tools can play an important role; some can make good decisions about plant operation states like expert humans do.”
Both fast and accurate
The neural network approach maintains remarkable accuracy despite its speed advantage, with an average heat absorption difference of just 0.3% compared to the traditional Forristall model. The research team demonstrated the model’s ability to accurately capture transient thermal behavior such as pressure drops and predict pump cavitation risks, validating their simulation using real-world data from the Solana CSP plant in Arizona.
In follow-on work, Wagner and Tuman have preliminary results showing that an AI-based control system negotiates the delicate balance between maximizing thermal efficiency and minimizing pumping power consumption. Through a parametric sweep of mass flow variability and total sector flow rate, the neural network model identified optimal operating conditions that maximize net electricity production. The simulation guides plant operators toward optimal control strategies for real-world operations by evaluating crucial trade-offs between maintaining high HTF outlet temperatures and managing energy losses in the system.
Based on the findings, workers in the power block or out in the solar field need to make some physical adjustments, working with the pumps, valves, and heat exchanger equipment. But most of the plant operation is computerized.
Urat Trough CSP plant in Mongolia ©Royaltech, China Nuclear
“At Solana in Arizona, for example, there are multiple operators in a control room,” Wagner explained.
“They’ll have half a dozen highly detailed screens in front of them, and then up on a display, a map of the temperatures of each collector or each loop in the field. They’re looking for places where the temperatures are not where they’re supposed to be, or you’re starting to see errors or warning markers. If they see that something’s not quite where it should be, like the temperature coming back from the field is too low or too high, just from sitting at the computer, they’ll modify the pump or the valve position speed and get the conditions back to where they want them to be. They can set the value there to increase pump speed from 50 to 52% or whatever, and it will respond.”
By enabling faster and more comprehensive simulation, this new neural network could help address one of Trough-type CSP’s key challenges: maximizing operational efficiency while reducing maintenance costs. For plant operators, as a bonus, this simulation also speeds up training new operators.
“These plants tend to be way out in the middle of the desert,” he observed.
“So one of the motivating factors is this higher turnover due to the typically remote locations of CSP plants. We are trying to help them quickly train and up-skill their operators. So we think this will have a lot of value for improving how people train on CSP systems.”
Solar energy record: Mongolian CSP generated round the clock
ACWA Power’s Bokpoort Sets New Production Record as South African Load-shedding Persists
