Explainable Deep Learning Models for Long-Term Temperature and Rainfall Forecasting

Enhanced Short-term Temperature Prediction Through ARIMA-LSTM Hybrid Model

Short-term temperature prediction plays a vital role in agriculture, water resource management, energy planning and climate change monitoring. Nevertheless, traditional statistical methods do not account for the nonlinearities in weather patterns, whereas deep learning models may not consider the linear time series characteristics. To solve this problem, the present study proposes a hybrid Auto-Regressive Integrated Moving Average-Long Short Term Memory (ARIMA-LSTM) model for short-term temperature prediction. The ARIMA model will be used in the proposed framework to identify the linear and seasonal components of temperature patterns. The errors produced by the ARIMA model will be further analyzed by means of an LSTM neural network that would model the non-linear time series behavior. The model is tested on the Tamil Nadu and Puducherry Weather Dataset (2022), consisting of roughly 1.8 million weather data sets, including variables like temperature of the air, humidity level, wind speed, and pressure. The evaluation criteria include mean squared error (MSE), root mean squared error (RMSE), mean absolute error (MAE), and coefficient of determination (R2). The experimental findings reveal that the developed model performs better in terms of accuracy than traditional deep learning models, thus providing an effective and reliable tool for predicting temperature in short intervals.

A CNN–LSTM Deep Learning Framework for Multivariate Temperature Forecasting and Extreme Weather Event Analysis

Throughout history weather predication has been a powerful and necessary tool in a number of activities that have had a major influence on human development and survival such as climate monitoring, agricultural planning and disaster management. Unlike other variables, temperature fluctuations are a major challenge for prediction due to the fact that they are nonlinear and extremely dynamic. This research paper introduces deep learning (DL) architectures for multivariate temperature forecasting using past weather data of single cities in India as samples. The DL methods such-as LSTM, GRU, Hybrid CNN-LSTM and Attention-based model were first outlined and then experimented. The meteorological variables that were considered include humidity, precipitation, wind speed and cloud cover. The quality of forecasting is quantified by the means of MAE, RMSE, MAPE and R2. Experimental results reveal that the DL methods achieve significantly better forecasting of temperature as compared to the conventional ARIMA model. Out of the deep learning models which were experimented with the CNN-LSTM model gave the best performance and has the following results: MAE (287.37), RMSE (401.60), MAPE (7.35%) and R2 (0.893). Besides that, CNN-LSTM model not only performed better in normal conditions but also in extreme temperatures and was able to achieve R2 of 0.923, which underscores its capacities to capture complex weather changes. This paper provides strong evidence of the value of combining DL methods for weather prediction. In the future, efforts will be made to enhance the accuracy of prediction by using transformer-based time-series models and incorporating larger spatiotemporal climate data.

Deep Learning Based Climate Forecasting Model Using LSTM (Long Short-Term Memory) of Kathmandu Valley

Deep learning’s significant potential as a scalable and dependable method for regional climate forecasting, giving it an advantage in the Kathmandu Valley’s climate change adaptation policy-making process is supported.

Explainable Deep Learning for Multi-Step Meteorological Forecasting in Saudi Arabia: A Foundation for Air Quality Prediction

Accurate and interpretable forecasting of meteorological variables is essential for environmental monitoring and for the development of reliable decision-support systems. This study proposes an explainable multi-step deep learning framework for forecasting daily mean air temperature in central Saudi Arabia. The dataset comprises daily observations collected from 2020 to 2024, including air temperature, atmospheric pressure, relative humidity, and rainfall. Historical measurements from a 30-day lookback window were used to generate direct forecasts for 7-day and 30-day horizons. Four forecasting approaches were evaluated: Seasonal Autoregressive Integrated Moving Average (SARIMA), Long Short-Term Memory (LSTM), one-dimensional Convolutional Neural Network (CNN1D), and Transformer. Forecasting performance was assessed using Mean Absolute Error, Root Mean Square Error, and Mean Absolute Percentage Error, together with Taylor diagrams, residual distributions, and observed-versus-predicted analyses. The LSTM model achieved the highest predictive accuracy, obtaining MAE and RMSE values of 2.139 °C and 2.953 °C, respectively, for the 7-day horizon, and 2.345 °C and 3.134 °C for the 30-day horizon. To investigate model behavior, seven complementary explainable artificial intelligence methods were applied, including Integrated Gradients, Grad-CAM, SHAP, LIME, permutation importance, occlusion sensitivity, and saliency maps. These methods were selected to provide global feature-level, local prediction-level, and temporal explanations. The results show that recent air-temperature observations dominate short-term forecasts, whereas atmospheric pressure and relative humidity exhibit greater relative influence at the longer forecasting horizon. Rainfall contributes less consistently because of its sparse distribution within the study region. Overall, the proposed framework combines multi-horizon forecasting with comprehensive interpretability, providing a transparent approach for meteorological prediction and a methodological foundation for future environmental forecasting systems that integrate meteorological and pollutant observations.

Long-term trend analysis and seasonal forecasting of temperature and rainfall in Sundarban Biosphere Reserve, India using non-parametric and hybrid deep learning approaches

Explainable hourly global solar radiation forecasting using a CNN-BiLSTM model with temperature scaled softmax attention.

Accurate global solar radiation (GSR) forecasting is vital for smart grids and resilient energy systems. However, the nonlinear and non-stationary nature of meteorological drivers challenges conventional approaches. This study proposes a lightweight, explainable hybrid deep learning architecture, CNN-BiLSTM-STAM, which integrates convolutional layers for inter-feature pattern extraction, bidirectional long short-term memory (BiLSTM) networks for temporal dependency learning, and a softmax temperature attention mechanism (STAM) with a learnable temperature to adaptively sharpen or smooth attention over time under varying weather conditions. The model is trained and evaluated using an hourly multivariate meteorological dataset collected at Alparslan Türkeş Science and Technology University in Sarıçam, Adana, Türkiye. Results show that CNN-BiLSTM-STAM outperforms classical machine-learning models, baseline deep networks, and recent attention-based variants for short and mid-term GSR forecasting, achieving RMSE = 66.62 W/m[Formula: see text], MAE = 43.29 W/m[Formula: see text], [Formula: see text] = 0.9381, and Pearson r = 0.9715. Additional validation on two independent public meteorological datasets, with the same architecture trained and evaluated separately on each dataset, yielded [Formula: see text] values of 0.9188 and 0.9537 and Pearson correlation coefficients of 0.9597 and 0.9768, supporting the external reproducibility of the proposed framework. To support operational trust, Shapley additive explanations (SHAP) are used to quantify feature importance and provide interpretable insights into seasonal and short-term atmospheric influences. With a compact footprint and strong predictive accuracy, the proposed framework supports deployable GSR forecasting for PV operation and scheduling, real-time energy management, and IoT-enabled monitoring systems.