Predicting Energy Consumption

Problem statement

The energy consumption patterns of customers are influenced by various factors, including weather conditions, as well as individual usage habits. To optimize energy distribution and reduce costs, electricity providers need accurate predictions of consumption. This requires advanced analytical tools that can process historical data and real-time weather information.

Additionally, the system of energy prices was about to change and the consumers had to estimate the maximum consumption for each of the 6 cost segments in a day. If consumption exceeded the estimated maximum in any segment, the consumer would face significantly higher costs. Therefore, accurate predictions were essential for cost-effective energy management.

System description

We first analyzed historical energy consumption data alongside weather data to identify patterns and correlations. Ideally, the customers could be clustered into groups with similar consumption patterns throughout the day, allowing a smaller set of models to be trained for each cluster. However, it turned out that smaller set of models predict the joint consumption of the cluster well, but fail to predict the consumption of individual customers on a 15-minute daily interval accurately. The best results were achieved by having a separate model for each customer which posed a challenge of training and running daily such a large set of models that exceeded 120 000 customers.

We overcame the challenge by identifying for which customers we could use a simpler model that required less training time and resources. For the rest of the customers, we used Long short-term memory (LSTM) neural networks implemented in Python using PyTorch. The models were trained on historical consumption data and weather forecasts to predict energy consumption for each 15-minute interval throughout the day.

Besides accounting for the weather forecast, the models also took into consideration the day of the week, holidays, time of the year, as these factors significantly influence consumption patterns as well.

Outcome

The system was successfully deployed on the customer’s infrastructure and is run daily to provide predictions for the next week (limited by the length of quality weather predictions) as well as a yearly forecast. The predictions are used by the electricity provider to optimize energy distribution, manage load, and help customers avoid excessive costs by accurately estimating their maximum consumption for each cost segment.