Abstract:
Air Quality Index (AQI) forecasting is essential for public health and environmental policy. While accurate, traditional numerical weather prediction (NWP) models demand vast computational resources. Recent trends have favoured deep learning data- driven approaches like Long Short-Term Memory (LSTM) networks, which model temporal dependencies in pollutant data. However, these methods typically require large datasets and a significant number of trainable parameters, making them inefficient and potentially overparameterised. In contrast, hybrid quantum-classical (HQC) models have emerged, incorporating parameterised quantum circuits (PQCs) as core learning components while deferring operations such as memory retention and updates to classical sub-networks. Although these approaches reduce parameter counts and introduce some quantum mechanical properties, the presence of classical components means that quantum features like entanglement and superposition are not maintained throughout the network. In this work, we explore a novel direction by investigating a fully quantum Long Short-Term Memory (QLSTM) architecture, a recurrent neural network constructed entirely from PQC layers. Unlike HQC models, our architecture performs all operations, including memory retention and updates, purely in the quantum domain. The model is implemented and simulated using PennyLane, and its learning and generalisation performance is benchmarked against classical and hybrid LSTM baselines. Despite hardware limitations and noiseless simulation constraints, QLSTM demonstrates convergence and learning capability. Although it does not outperform its classical or hybrid counterparts in short-term AQI forecasting, it establishes the feasibility of full quantum recurrence in temporal modelling. We also explore optimisation strategies such as RMSProp, based on its successful use in previous hybrid models. Our findings suggest that fully quantum recurrent networks are viable on near-term quantum simulators with limited qubit budgets. This work opens avenues for future research in circuit compression, noise-aware training, and theoretical analysis of quantum memory mechanisms. The proposed QLSTM thus lays the foundational groundwork for deeper integration of quantum computing into time series learning tasks.
