Electricity suppliers are working hard to deliver the right amount of power to the grid in the right time. The amount should correspond to the expected consumption of their customers. A lot is at stake - though the regulators would buy out or supply the excess/missing power in most cases, incorrect estimates of the consumption by energy suppliers are penalized.
Electricity suppliers, such as V-Elektra, therefore need accurate predictive models to guide their market operations.
Cognexa took on the challenge to build a new models, which would improve the accuracy of prediction over the previously used models. After just 2 weeks of intense sprint, our solution was deployed and is continually improving its performance since then - both thanks to its self-learning and our upgrades. We also ensure availability and quality of the predictions.
The model estimates the expected consumption of the company’s clients and amounts of available electricity for the next day in hourly granularity. We fuel the models with variety of data sources, including the weather records and forecasts, previous values of consumption and production, prices and various data published by the regulator.
At the end of Q1 2018, the core model was an ensemble of several neural nets. During the rapid prototyping phase, as well as after the deployment, we experimented with usage of various data sources, sliding window sizes, model families and regularizations. The solution was developed using cxflow and Tensorflow.
The neural nets proved to be a more accurate option, when compared to the more conventional time series predictive models, such as SARIMA. This is mostly due to our innovative approach to regularization, which makes the models plastically adaptable to the changing environment and accurate in prediction at the same time.
The accurate consumption and production estimation is a solid basis for all sorts of decisions that an energy supplier has to make. After three months in production, the relevancy of predictions is high and customer is satisfied.
Time to production of just 2 weeks is extremely short, even in Cognexa. However, our teams have exceptional skill in rapid prototyping and we managed to deliver a good performing model from the day 1.
Fair and motivating contract terms and good relationship between Cognexa and V-elektra is benefiting both sides. Working on the models for V-elektra have become very popular among our team and we work hard on improving it. There is even an internal competition among our data scientists to reach the best performance on the copy of production data. This not only ignites creativity and fosters our research, but also makes our work more fun.
Electricity consumption and production predictions on day to day market
2 weeks to deployment, continuous improvement since then
Python, Tensorflow, CXFlow
Our model is the most accurate in the company’s portfolio