Dapesco, part of Metron Group, is a software company which offers an energy management system (EMS) for global organizations. This includes international retailers (such as L’Oreal and Carrefour) as well as more energy-intensive tertiary sites including hospitals and airports. This integrated solution gives clients a more global understanding of their energy consumption and carbon footprint data.
Dapesco wanted to enhance its existing software platform, by adding predictive capabilities to improve energy efficiency, detect anomalies in electricity usage, and provide actionable insights for retail clients across multiple stores.
The Complexity of Achieving High Accuracy in Predictive Modelling
Dapesco faced several challenges in building an effective predictive model due to the lack of a clean set of data:
-
Unpredictive variables: the energy consumption patterns were influenced by multiple factors such as store hours, weather, and seasonal changes. This variability made it difficult to build a reliable forecasting model.
-
Noisy data: the raw data included irrelevant and noisy variables, such as dummy variables and constant values, which made the useless in some cases for building the model.
-
Lack of clear context for some variables: many variables in the dataset, such as solar intensity and day of the week, lacked clear context, making it difficult to determine how these variables impacted energy consumption.
-
Irregular data patterns: the data exhibited sudden shifts and high-frequency noise, which needed to be cleaned and processed for effective forecasting
Solution: AI-Driven Forecasting & Anomaly Detection
Faktion partnered up with Dapesco for the development of a predictive model tailored to Dapesco’s needs. We focused on developing a model that could accurately forecast electricity usage and detect anomalies in energy consumption. By integrating various data sources and features, such as store opening hours, outdoor temperature, and solar intensity, we built a model that could capture the nuances of energy consumption across different retail environments. The model used advanced regression techniques to provide highly accurate energy usage forecasts, enabling Dapesco to detect inefficiencies or unusual consumption patterns promptly.
Our Approach
We followed a detailed and iterative approach to ensure the model was tailored to Dapesco’s needs:
-
Data exploration: we began by analysing the provided dataset to understand key factors affecting electricity consumption, like temperature, solar intensity, and day of the week
-
Feature engineering: we created additional features and adjusted the data to account for specific store behaviours, like energy usage during restocking
-
Handling data irregularities: we filtered out unusual patterns that could skew the model’s predictions
-
Iterative model development: we used Lasso Regression, which allowed the model to focus on the most relevant features while reducing the impact of irrelevant ones
-
Model evaluation and fine-tuning: we fine-tuned the model by iterating on the feature selection and model parameters to ensure optimal performance
Outcome: Precise Predictions and Real-World Impact
By incorporating key features such as store hours, holidays, and seasonal changes, the AI system explained 93,4% of the variance in electricity usage and hence, it was able to generate precise forecasts tailored to each store’s consumption patterns. Also, the model was designed to be scalable, with the ability to apply the same methodology to multiple stores and adjust for each store’s unique characteristics
This will bring two benefits for the platform’s end-users:
-
Cost saving: The insights gained from the model allowed Dapesco’s retail clients to take action to optimise energy usage, potentially leading to significant cost savings over time
-
Real-time energy monitoring: In addition, the model’s ability to detect deviations (anomalies) from expected usage patterns allowed Dapesco to identify potential inefficiencies or issues in real-time.
Looking ahead, the success of this project has set the foundation for future development, including the integration of additional weather data and the possibility of clustering similar stores for better benchmarking and model improvements
