Autonomous Agents can boost Wind Energy value

Game-Changing Renewables

This will be the first post covering technical depth in using Decision-Making agents for Wind Farms. We decided to cover the use of Wind Farm due to their price plummeting for the recent 10 years, which made it considerably, a feasible alternative for fossil-fueled of combined-cycle power plants. Furthermore, the baseline for our service works tremendously well with an arrangement like the one found in wind farms: individual components of single mills working to produce together, to increase the fitness of the collective or the wind farm as a whole; similarly to an ant colony.

Wind Farm: The Collective & The Unit

The main problem with wind farms, a serious one, is weather prediction. Weather forecast is essential as a value for such business due to the allocation of energy-generating assets beforehand e.g. a wind energy supplier should publish how much energy he can provide to the grid at a time not less than 24 hours; this can be challenging from an allocation of assets perspective when uncertainty, that is, weather prediction, comes into play.

In my personal view, I have been a long-time skeptic about predictions that cost a lot once not correct and further, I don’t think an ant or bee colony is equipped with forecasting tools that can increase their fitness in the long run.

Therefore, let’s explore a different process than the weather forecast and focus a bit on the allocation of resources.

Forget about Weather Forecast, again

If we want to go with the logic of an ant colony, by the collective dynamics we can agree that we have an idea about how the individual works however, we don’t have a great idea about the collective how it works. Nassim Nicolas Taleb mentioned many examples in his book Antifragile about the collective and the unit in quite a hyperactive way and he said that we know how a neuron works but we don’t a slight godd*mn idea how the 300-neuron system of the Caenorhabditis elegans works to this day. Similarly, we know how an individual seller or buyer works but, unfortunately, we have no knowledge at all about the market behavior and its dynamics. Nonetheless, it has been proven that a market behavior won’t change whether its main constituents become more idiots.

Systems increase their fitness by simply removing bad components. They work by removing, not by adding.

Increasing Payoff Value not Prediction Accuracy

So we follow up with this central idea here and we will consider that we know how a single wind mill works but not the collective wind farm. And next, we want to apply a multi-agent algorithm couple with RL and some Non-Gaussian distributions to account for fatter events. Then the idea will be to disregard completely the weather prediction and let the collective decide the best allocation of assets at any given time.

Now for the skeptics who can’t operate their minds outside the illusion of safe prediction we can simply tell them to consider an operation of wind farms with less mistakes; sounds not a bad idea as an alternative. Well if we look at it in this way, as an asset manager for a wind farm, how can I make my judgment better in the long run for better allocation of power delivery? Obviously, I can look for the past data and check probabilistically what is the best base load that can keep me safe, at the lowest cost, for any volatility in demand and weather:

“Historically I was better off by operating 70% of my wind farm at high speed on a certain angle, and 20% at mid-speed at another angle and 10% etc…” So such ideas can be generated by a multi-agent decision-making system where it looks at historical data and checks for weaknesses or mistakes in the management or allocation of individual wind mills. The whole idea is to keep oneself safe at any decisive moment. The idea is very similar to options trading in financial markets: a cheap option betting on any market can be substantially cheap, but benefits a lot on long term volatility. In our case, we should strive for always having the cheapest base load allocation at any moment in time. The most important thing is that we don’t to offer an allocated-asset that will cost me a lot if the prediction was false.

Implementation

To go more in details on how to build an Autonomous Decision-Making Agent for our wind farm it suffices to simply build first our State/Action vector described previously in other posts and then proceed to configure the desired target. But first let’s re-examine of what we really want to achieve as a final objective. Previously we stressed on the importance of non-interventionism in complex domains prediction and here, in our wind farm issue. weather forecast is essential to our final outcome. But weather forecast is similar to earthquake prediction, maybe not at the same level of magnitude but still, forecasts based on weather can cause massive losses once not correct and for that we need to go for the other alternative: increasing our upside against the convexity of weather volatility. So basically, I can decrease my downside by simply analyzing past mistakes during the operation of my wind farm: mistakes are meant but the non-systematic ones, that are outside of our control. Mistakes can be simply the non-desired actions that were taken at a specific moment and could’ve been replaced with better action strategies. For example, I should’ve operated 70% of mind wind mills high speed, 20% mid-speed and 10% low speed during the 1st week of August, because I saw that this configuration in fact brought me a bigger payoff as into money value.

Adjusting my actions based on forecast and doing the right actions to increase my payoff value are not the same thing.

Actions come First. Everything else, after

Hence, the idea of our algorithm is to propose the right actions that increase our future payoff regardless of any forecast. Like governments impose residents to build homes under a specific standard for protection against earthquakes, we as well, need to have always in mind to provide the right amount of power that keep us safe at any specific moment. Medical insurance does the same thing: we pay a small premium (low cost action) to protect ourselves on the long-run. Now consider when an insurance strike on a very expensive surgery: you paid a small premium but you can have an expensive operation with that same premium (extreme profits). So the idea again, is to keep an acceptable amount of base operation of wind mills that can protect us eventually from any unexpected demand or market price fluctuation.

Case in Point: Alphas & Hedge Funds

The study above reminds me of Hedge Fund managers who are always looking for a certain “Alpha”. It turns out that “Alpha” isn’t meant by a favorable condition in the market but rather a favorable allocation of assets for investors at any given moment regardless of any market volatility. Recently, there has been many Hedge Funds turning to AI and RL and retrofitting their systems to find “Alphas” dynamically through complex computations or in other words, shuffling their invested capital more dynamically than the human way, where it is able in this way to catch and learn from many possibilities and many possible allocations of assets combinations.

First Principles, again & again

We might still need to understand what’s happening in the background but again, as we mentioned before, RL and Probabilities are just complex enough to be explained on a human scale and the only logic answer would be to stick to an algorithm that provides good long-term actions for a robust payoff. Nothing more, nothing less.

Before we conclude, the application of such solution can be done without any interference in the mechanical dynamics for the on-site wind mills and will be enough to be implemented with just the historical data collected: action and feedback data. Targets can be simply a maximizing objective function for the total profits or a cost minimization function.

Conclusion

To end this post, we should consider that in a world of complexity, it makes sense to implement algorithms that can achieve an ROI with not less than 15 to 20 percent. Google’s Deepmind did a lot recently for the AI & RL processes in the renewable sector and successfully produced a 20 percent increase in wind-farm output value. With the price of wind mills plummeting, the game is changing fast to say the least.