Is it going to rain tomorrow? Which stock should I buy? Which team is going to win? And of course – how do I see myself in five years? We all rely on predictions a lot! But… why? Why are we so interested in the future? And how are we even able to predict it without a crystal ball?
Predictions, if accurate enough, enable us to make smarter decisions before something happens. The key ingredient for making stunningly accurate predictions are previous experiences. You can think of these experiences as arrays of events – a temporal order in which events occur. And you have a natural talent for making day-to-day predictions just by recognizing patterns in which events occur! It’s easy for you to pinpoint events that seem to ‘cause’ others.
If we’re so good at predicting stuff, why can’t we easily predict everything? It’s because some things are hard or even impossible to predict. That’s why we rely on the power of computers to detect patterns we ourselves could never see, by taking many more features into account.
The brief: predict future energy consumption SmartCat was requested to figure out how much electrical energy something will have had spent… in the past in the future!!! Future past? Future perfect tense? You get the idea: predict consumption based on data that contains large sequences of numbers that really don’t make much sense. It’s basically impossible for a human to parse all that in her lifetime. Unless you use the power of machine learning.
We used a regression-based neural network to solve this problem
We presented it as a supervised learning problem. The neural network makes predictions by taking previous events into account – just like you would do. By feeding our network the data that describes the sequence of events that happened in the past, and bringing the data of the last event to the output of the network, we made it tune its weights that essentially describe the pattern in which these events occur. This network was later used to predict future consumption. The image above shows predicted values and true values, and we can see that these are highly correlated. Our neural network is able to make stunningly accurate predictions. No tea leaves necessary.
So how can an energy-providing company benefit from using machine learning like this? Results like these will assist in establishing the supply and demand equilibrium. Predictions can help a company in making informed decisions such as:
- setting a limit on how many customers they have,
- defining the price,
- handling predicted peak demand periods
- and many more.
And most importantly: load prediction helped us to save energy! By knowing what to expect, we were able to tune consumption profiles that we recommend to clients. You can read more about this in our previous blog about load prediction.
And that’s all there is to it! Easy as pie, right? 😉