# Redefining multi-target weather forecasting with a novel deep learning model: Hierarchical temporal convolutional long short-term memory with attention (HTC-LSTM-Attn) in Bangladesh

## TL;DR

A new deep learning model called HTC-LSTM-Attn is developed to forecast temperature and humidity in 24 cities in Bangladesh, showing strong performance for agricultural and disaster planning.

## Contribution

The novel HTC-LSTM-Attn model combines hierarchical temporal convolutions, bidirectional LSTMs, and attention for improved weather forecasting.

## Key findings

- The model achieves MAE of 0.8178°C for temperature and 2.4693% for humidity on a temporal test set.
- Performance remains robust on a spatial hold-out test set with MAE of 0.9587°C for temperature and 2.4796% for humidity.
- The model outperforms existing architectures like GRU, LSTM, and transformer-based models.

## Abstract

Bangladesh, being a country that relies on agriculture, has its agricultural sector contributing around 20% to the Gross Domestic Product (GDP) and providing employment opportunities for nearly 60% of the population. The study proposes a novel multi-step forecasting deep-learning framework designed to forecast the maximum temperature and humidity of 24 cities, called Hierarchical Temporal Convolutional Long Short-Term Memory with attention (HTC-LSTM-Attn). The model contains hierarchical temporal convolutions (HTC) for extracting multi-scale patterns, bidirectional LSTMs for sequential dependencies, and an attention mechanism for learning to weigh time steps differently, while parameters are optimized with Keras tuner. Using data from Bangladesh Agricultural Research Council (BARC) from 1961 to 2022 and rigorous preprocessing with seasonal features and lagged statistics, the model obtains temperature results for with a Mean Absolute Error (MAE) of 0.8178 °C, Root Mean Squared Error (RMSE) of 0.9718 °C, R-squared (R²) of 0.8527, and Mean Absolute Percentage Error (MAPE) of 2.8823% and humidity with MAE of 2.4693%, RMSE of 3.2442%, R² of 0.7228, and MAPE of 3.1757% values on a strict temporal test set (19 stations, 2016–2022). On a spatial hold-out test set (5 unseen stations, 2016–2022) performance remains robust, temperature with MAE of 0.9587 °C, RMSE of 1.1898 °C, R² of 0.8342, and MAPE of 3.1534% and humidity with MAE of 2.4796%, RMSE of 3.2119%, R² of 0.6561, and MAPE of 3.2045% consistently surpassing performance measures of models Gated Recurrent Unit (GRU), Long Short-Term Memory (LSTM), Convolutional Neural Network (CNN), Decision Tree Regression (DTR), Convolutional Neural Network – Long Short-Term Memory (CNN-LSTM) and latest transformer-based architectures (Autoformer, FEDformer, TimesNet, and Pyraformer). These results have shown robust spatial performance with the respective improvements needed in agricultural planning and disaster management in Bangladesh, in its vulnerable climate context.