ConForMiSt: A Multi-model Dual-phase Framework Utilizing Machine Learning for Carbon Footprint Prediction and Reinforcement Learning for Decision Optimization

Over the past decade, there has been a significant surge in harmful waste emissions of greenhouse gases, namely carbon dioxide, methane, and fluorinated gases, in the atmosphere. Two major categories of activities can be broadly identified as contributing to this condition. The first is the proliferation of worldwide industrial activity, as accounted for by industrial plants across all major continents. Second, human activity contributes to carbon emissions produced as a result of wide-ranging everyday activities that involve the use of electricity, transportation, food consumption, and other consumer-mindset-driven activities. This study focuses on the second category to build a dual-stage framework that will assess, evaluate, and recommend suitable mitigation measures to regulate usage patterns. The dual-stage approach is a novelty based on sound engineering principles. Carbon emission data gathered by the system were analyzed to detect footprint generation patterns using mathematical models. After the analysis, machine learning models selected from rigorous performance metrics (MAE, RMSE) were leveraged to predict the carbon footprint in the first stage. The second stage employs a reinforcement learning framework that captures several aspects of emissions in a ‘state’ and is used to analyze predictions and generate recommendations considering user preferences. The ability to absorb user goals for emission data is a strength. This unique finer engineering of state representation exemplifies experimental data that show minimal variation in state goal values within 2000 steps. A web application was developed to visualize various aspects, such as usage patterns and predictions. The user interface provides interventional, specific, and personalized recommendations. These aspects are then utilized to provide insights at the aggregated level in the context of a group of individuals, which is yet another strength of this framework. The extensibility of the proposed methodology for carbon emission mitigation for higher aggregated levels is demonstrated by an exemplar ‘location statistic’ radar chart in the context of the vehicle and electrical appliances categories.