We fed our hydropower turbine data to an AI machine. What happened next was astonishing.

My life has revolved around the art and science of hydropower for the past four decades. I've seen and steered innovations in the sector, from the most minute adjustments in turbine design to expansive projects such as the XFLEX HYDRO project. Recently, my collaborators at the University of Catalonia (UPC) and our EDF hydro team ventured into an experiment that breathed new life into my established world view of hydropower systems.

For years, our understanding of turbines rested on carefully constructed, precise, physical equations. Imagine an elaborate Lego castle, each brick meticulously placed based on detailed plans, and every tower engineered for optimal strength. This intricate world of bricks and instructions mirrors our traditional method of predicting turbine performance.

The XFLEX HYDRO project, however, compelled us to push the boundaries. We wanted to measure the impact of integrating a small battery into the Vogelgrun run-of-river plant's hydropower unit, aiming to reduce turbine wear and tear by a factor of ten. But how could we compare the effects of this battery-hybridized unit with a non-hybridized version without a suitable point of reference?

The non-hybrid version of our turbine did not exist anymore: the battery had been installed. We had to come up with a way to simulate how the non-hybrid turbine would have behaved over the same period of time.

This was where our partners at UPC proposed an audacious idea - why not feed our turbine data to an AI? The suggestion was the equivalent of tossing our meticulous Lego instructions to a curious novice, albeit one with an astonishing capacity for learning.

Sceptical, I agreed, and we began to feed our AI an entire set of data representing two months of turbine operation. Every guide vane and blade opening, unit power, water head and grid frequency information went into this training set. The AI was to build its own predictive model, much like the novice constructing our castle without the instructions.

"Every guide vane and blade opening, unit power, water head and grid frequency information went into this training set"

What followed was a revelation that took many of us by surprise. The AI predictions were eerily accurate, deviating by less than 2% from our traditional equation-based predictions. It was a paradigm shift akin to watching the novice not only build the castle but add unforeseen enhancements, all while maintaining the structural integrity.

Despite this success, does AI spell the end for our time-honoured physical equations? Not quite. While AI offers fascinating capabilities, changes in turbine operation require the retraining of our AI model. Meanwhile, our equations can adjust more readily. Thus, the old and the new, equations and AI, must co-exist, each complementing the other.

One of the important lessons from our project was the need for diverse approaches. Yes, the AI was able to predict turbine performance with remarkable accuracy. However, it was the synergy of multiple methods - including the use of physical equations, the neural network, site measurements, and the DiOMera software by Andritz - that painted a complete and reliable picture.

An AI model's predictions are contingent upon the quality of the data it's trained on, and in our case, the data represented a specific operational state of our turbines. When we adjust our turbines' settings, our equations can adapt and continue providing insights, whereas our AI model needs relearning. This is not a weakness per se, but rather a distinction in the application and adaptability of these tools.

So would AI work best? I can imagine there would be instances when equations become exceedingly complicated due to numerous variables or become hard to formulate for novel situations. That's where the AI, with its data-centric approach, might provide a compelling alternative.

"The comparison of our turbine's behaviour in hybrid and non-hybrid cases was just a start."

The comparison of our turbine's behaviour in hybrid and non-hybrid cases was just a start. We now see opportunities to apply this AI-centric approach to areas where conventional wisdom may struggle, for example assessing the overall health of a power plant, civil and electrical infrastructure included - similar to how AI is revolutionizing the medical field.

The integration of AI in hydropower engineering is not a wholesale replacement of our established methods. Instead, it's an augmentation that broadens our arsenal of tools and widens our range of possibilities. The confluence of our collective expertise and AI's ability to learn and predict will keep driving the field of hydropower forward. It's not just about turbines anymore - it's about pioneering a future where tradition and innovation converge in the most unexpected ways.

Forty years in the hydropower industry, and it still manages to surprise me. The AI's unexpected 2% accuracy felt like a nod from the future, an affirmation that our industry is far from reaching its full potential. The tools have changed, but our pursuit remains the same – to turn the might of water into a resource that lights up our world. After all, isn't that the beauty of being in this industry for so long? The river keeps flowing, but every now and then, it shows you a better way to harness its power.

‍