Environmental Data Exploration and Modelling Using Extreme Learning Machine

Extreme Learning Machine (ELM) is a fast and efficient learning algorithm designed for single-layer feedforward neural networks. It stands out by randomly initializing input weights and biases, which remain fixed during training, and learning the output weights using a closed-form solution. This approach eliminates the need for iterative optimization, significantly accelerating the training process. ELM is known for its generalization performance and versatility. Kernelized ELM  enhance its capability to model complex nonlinear systems. However, achieving optimal performance requires careful tuning of hyperparameters, such as the number of hidden neurons and the regularization parameter.

 

ELM has been widely applied in environmental risk and natural hazard assessments, climate and meteorological modelling, hydrology, renewable energy analysis  and time series forecasting. Recent advancements have extended the standard ELM model to include multilayer architectures, deep learning methodologies, unsupervised learning, and multiple kernel ELMs, broadening its applicability to more challenging  and diverse problems.

 

This research investigates the application of ELM for intelligent environmental data exploration and modelling. The study focuses on addressing problems in spatial and spatio-temporal data exploration, analysis, and modelling, including feature engineering and selection, multi-scale analysis, data normalization and anisotropy, nonlinearity, multivariate analysis and uncertainty quantification. 

 

The quality of ELM-based modelling is assessed through the examination of unexplained variability in  data and a comprehensive analysis of residuals. Various ELM configurations are applied throughout all phases of the research, enabling a flexible approach. Due to its computational efficiency, ELM facilitates numerous simulations and experiments, providing deeper insights into the data and the resulting models. Both simulated and real-world environmental datasets, including pollution, precipitation, and permafrost data, are utilized. Finally, the performance of ELM is compared with other machine learning algorithms in order to evaluate its effectiveness and reliability.