Deep Learning Methods to Enhance IoT-based Air Pollution Forecasting

Abstract

The research endeavor on utilizing deep learning for air pollution forecasting is escalating, driven by the availability of Internet of Things (IoT) based devices, the advent of deep learning, statistical modeling, and growing emphasis on climate change. An insight into the challenges in predicting pollution levels necessitates development and adoption of innovative approaches based on deep learning models. This thesis aims to develop new deep learning algorithm models enabling early information pertaining to pollutant concentration detection in specific locations. However, task of pollution forecasting is fraught with challenges due to the complex multivariate pollutants, weather conditions, and intricate spatial-temporal dynamics of nearby areas. To address the aforementioned challenges, this research proposes an array of strategies leveraging IoT devices and deep learning algorithms to enhance air quality monitoring, aligning with sustainable development goals related to clean air, sustainable cities, and climate action. Novel deep learning frameworks showcased in this thesis aim to streamline the computational processes while effectively extracting spatiotemporal characteristics, resulting in improved pollution prediction accuracy. Use of optimized algorithms, hyperparameter tuning, and application of computational methods to real-time environment issues to solve complex air quality prediction problems enriches the contribution of this research. Firstly, an attention-based AQI forecasting model has been developed with a combination of bidirectional-gated recurrent unit (Bi-GRU) and bidirectional-long short term memory model (Bi-LSTM) networks. A gate attention mechanism and a denoising mechanism are incorporated into this model in order to minimize error rates, optimize attention weights, and improve prediction accuracy. It has a linear activation function, a dropout layer to create a 1D array, and a time-distributed layer with sigmoid and (rectified linear unit) ReLU activation functions. With the Adam optimizer and mean square loss function, optimization is accomplished. Secondly, a novel Dense Residual GRU-CNN Network with Attention Mechanism (DRAGCN) is developed, featuring Global Max Pooling and a 1D convolutional layer. This model performs well in both univariate and multivariate scenarios, preventing overfitting and disappearing gradients. Performance is enhanced, and training is stabilized through normalization using min-max scaling. It combines GRU’s gating mechanism with convolutional layers, Adam optimizer, and ReLU activation to capture long-term dependencies in sequential data, results in a non-negative scalar output to represent expected pollution levels. Furthermore, a deep learning-based metaheuristic optimization algorithm, the Xavier Reptile Switan-h-based LSTM model (XRSTH-LSTM), has been developed, which expounds the versatility of different frameworks used for air pollution prediction. It addresses the vanishing gradient problem by combining the Xavier Reptile Search Algorithm (XRSA) and LSTM networks with the Swish-Tanh (Switan-h) activation function. The dataset was expanded to evaluate the robustness of the model, and normalization was applied using the Min-Max method along with hidden layer optimization. Finally, a step forward has been taken with the successful implementation of a hardware design and product prototype application using a three-layered IoT architecture. This includes sensor calibration, PCB design, and GUI implementation, along with an Android application featuring alert notifications. The system utilizes Raspberry Pi for real-time monitoring of harmful pollutants, directing sensor data to a cloud database like Firebase. This IoT-based system can be practically applied and tailored further according to specific needs in air quality monitoring. The developed models were evaluated, and a comparative analysis was carried out based on prediction parameters to provide for timely decision-making and preemptive action. The proposed approaches in this research work collectively aim to improve the accuracy and efficacy of air pollution forecasting. Through rigorous testing, along with the comparative results obtained from the evaluation, these models demonstrate their efficiency and potential as valuable tools for decision-making processes in policy-making, environment, and biodiversity domains.