Applications for AI tools and machine learning, a case study with Kaleva Media

GOALS

• Assist Kaleva Media in saving energy, reducing costs, and decreasing carbon emissions during the printing process.
• Explore the potential of neural networks for projects related to energy consumption management.
• Enhance knowledge about utilizing ChatGPT as a tool in neural network development.

LIMITATIONS

• Inability to introduce new hardware.
• Rescheduling printing times to periods with cheaper electricity rates is not feasible due to strict printing timetables.

SKILLS AND TECHNOLOGIES

• ChatGPT
• Neural network
• Machine learning
• Python

Using machine learning to detect anomalies from energy consumption data

We combined the approaches of our software specialists with suggestions from ChatGPT and interviews with customer representatives. We discovered that Kaleva Media had an extensive record of their past energy consumption statistics. This meant we had tens of millions of rows of energy consumption data to start with.

This gave us the idea to use machine learning to detect anomalies in their energy consumption data, which might be used to flag faulty machinery before situations become dangerous. This necessitated the creation of Python scripts to gather and analyze the data.

Training neural networks sets requirements for data quality and quantity

We proceeded to train the neural network on how different variables and signals should be weighted. For anomaly detection, we opted for a type of neural network known as a 'Recurrent Neural Network' (RNN). This specialized neural network is adept at identifying patterns in data that involve a time dimension. Thus, in our context, the RNN would recognize if a specific kind of energy consumption typically followed another. This approach allowed us to develop a prediction model that outlined what “normal” energy consumption looked like, facilitating the identification of deviations from the norm. Now, we possessed an RNN equipped to discern patterns in the data. Our subsequent challenge was pinpointing the optimal parameters for the neural network to minimize the discrepancy between the predicted and actual data, ensuring anomalies were detected without mistakenly flagging standard events as irregular.

Data cleansing is paramount, perhaps the most critical step, when designing a neural network. If the data isn't meticulously cleaned, the resultant model would be ineffectual. Our experimentation affirmed that neural networks can be employed to spot discrepancies in energy consumption data. However, the data should be sufficiently detailed and span an adequate duration. This insight is invaluable for companies contemplating the adoption of such machine learning algorithms. The recorded observational data should be as comprehensive as possible, emphasizing that the present is an opportune moment to commence data accumulation.

Read the whole study including step by step report of the case >>

To try out what this new AI can be used for, our team of three is ready to help – the author of this article, Tapani Mattila, and Jussi Puhakka.