An Integrated Fuzzy Neural Network with Auto-Encoder for Time-Series Forecasting to Predict Air Quality

Lately, in the domain of wide-ranging time-series forecasting, there has been a significant increase in various architectures based on recurrent neural networks (RNNs) including the gated recurrent unit (GRU), long short-term memory (LSTM), recurrent highway network (RHN), and sequence-to-sequence (seq2seq). These models have proved remarkable capabilities in handling complicated time-series datasets. However, these approaches still face certain limitations. One major drawback of RNN-based architectures is their sensitivity to data noise when processing dynamic information from time-series data. These challenges arise during learning temporal feature representations and can lead to a decline in the performance of the fine-tuning process for forecasting tasks. To tackle this obstacle, our study proposes an innovative model named “FuzzAE4PAQ” for time-series forecasting, specifically applied to predicting air quality using real-world datasets. Our proposed FuzzAE4PAQ model combines fuzzy neural networks and auto-encoder architecture to downgrade the impact of feature noise during the learning process. Our proposed model demonstrates its effectiveness and outperformance when compared against existing state-of-the-art time-series forecasting benchmarks through experiments on real-world air quality datasets.