Predicting Next-Day Consumption for Aneo

Motivation

Power producers, power suppliers, and consumers participating in the power market have a complex task that requires domain knowledge, timing, and good instinct. However, power producers use information from production forecasting, while power suppliers and end-users use consumption forecasting. The two forecasts require different data and solve separate problems in the power market. As the team has acquired data on customer consumption, it is more relevant to look at how end-users or power suppliers could use the forecasts.

For Aneo, modeling a forecasting tool and selling forecasts are great sources of income. Smaller power suppliers may not have the resources to invest in forecasting themselves and rely on external expertise to gain information. Aneo can do the analysis for multiple customers with their data science team, providing accurate, high-quality forecasts. Considering the importance and value of the forecast information, Aneo most likely has long-term, consistent revenue as long as the forecasts yield accurate information for their customers.

The problem to solve

Given the motivation for this paper, the problem to solve is the following: How can we reliably predict tomorrow's energy consumption, using a data-driven approach? A successful forecast creates value for all parts, including the end-users, and reduces the uncertainties around power production and buying power. As stated by the US Environmental Protection Agency, all electricity has an environmental impact on nature, and efficient electricity helps reduce the impact. Thus, avoiding overproduction and loss of power through efficient energy management is crucial for sustainable energy consumption. The perfect forecast is (nearly) impossible to create, and, thus, there will always be room for improvement. Our team aims to improve the energy consumption forecast, comparing it with a basic baseline.

From a machine-learning perspetive, this is a time-series prediction problem. We used linear regression as our model.

Method and Analysis

Describing the dataset

The dataset consists of four main features: time, location, consumption, and temperature. The time feature indicates the moment data was recorded. Location is a nominal categorical feature representing the largest city in the price area, with six possible values. The consumption feature measures the average hourly energy consumption in megawatts (MW) for each location. Lastly, the temperature feature records the hourly weather forecast in Celsius for each location. The dataset, sourced from an anonymous power supplier, reflects the energy consumption of the supplier's customers in different geographic regions.

Preprocessing the data

Preprocessing involved handling missing values, particularly for the Helsingfors location, by applying a forward-fill method based on weekly consumption patterns. The process ensured the dataset was clean and suitable for modeling.

Feature engineering

Initial feature engineering included extracting time-based features (date, holiday status, weekday, hour of day) and temperature differences. These were used to capture daily and seasonal patterns in energy consumption. One-hot encoding was applied to the location feature to convert it into a numerical format. Iterative improvements involved analyzing the impact of holidays on consumption, introducing new features like holiday_season_winter, and adjusting parameters like the alpha value in exponential smoothing to better balance historical and recent data.

We found that consumption spikes during winter holidays, seemingly due to the cold weather and the holidays. This is illustrated in the image below.

We also found that consumption is generally lower on weekdays, with peaks in the morning and evening hours, and higher during weekends. This is illustrated in the image below.

Visualization of findings after feature engineering

Evaluating the model

Defining performance metrics

The primary metrics include Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE). These metrics were chosen for their ability to effectively measure forecast accuracy and error magnitude. MAE offers a straightforward error average, RMSE gives more weight to larger errors, and MAPE provides a relative error perspective, making these metrics comprehensive for assessing the model's performance.

Conclusion

The rest of the report is really the business-side of the case, not the technical implementation. I will not include it here, as it is not relevant for this portfolio.