CO(2) concentration forecasting in smart cities using a hybrid ARIMA–TFT model on multivariate time series IoT data

Carbon Dioxide (CO[Formula: see text] ) is a significant contributor to greenhouse gas emissions and one of the main drivers behind global warming and climate change. In spite of the global economic slowdown due to the COVID-19 pandemic, the global average atmospheric CO[Formula: see text] concentration reached a new record high in 2020 with its year-on-year increase being the fifth highest annual increase in 63 years, according to the National Oceanic and Atmospheric Administration. Furthermore, the years 2020 and 2019 were respectively the second and third warmest, while the decade 2010–2019 was the warmest decade ever recorded. In an attempt to curb this climate emergency, many countries and organizations globally have adopted ambitious goals and announced plans to help dramatically reduce CO[Formula: see text] emissions. As part of these plans, various innovative smart city projects are being developed, focusing on implementing Internet of Things (IoT) technologies. By collecting sensor-based data, such technologies aim towards automating data-driven decision-making around carbon emission management and reduction. In this work, a hybrid machine learning system, aimed at forecasting CO[Formula: see text] concentration levels in a smart city environment was developed using a multivariate time series dataset containing IoT sensor measurements of CO[Formula: see text] , as well as various environmental factors, taken at every second. The proposed system demonstrated superior performance to similar methods, while also maintaining a high degree of interpretability. More specifically, the approach was empirically compared against other similar approaches in several scenarios and use cases, thus also offering more insight into the predictive capabilities of such state-of-the-art systems. For this comparison, both traditional time series and deep learning approaches were employed, including the current state-of-the-art architectures, such as attention-based, transformer networks. Results demonstrated that, when measured across various settings and metrics, including three different forecasting horizons, the hybrid solution achieved the best overall results, and in some cases, the difference in performance was statistically significant. At the same time, insights from the system’s inner workings were extracted, shedding light on the reasoning behind the model’s predictions and the factors that contribute to them, thus showcasing its transparency. Lastly, throughout the experiments, deep learning approaches illustrated their ability to better handle the multivariate nature of the dataset and in general tended to outperform the traditional time series methods, especially for longer forecasting horizons.