Forecasting temperature and rainfall using deep learning for the challenging climates of Northern India

Introduction

Jammu & Kashmir (J&K) and Ladakh, located in the western Himalayas, are India’s northernmost union territories characterized by highly variable and extreme weather. These unpredictable climatic conditions significantly affect agriculture, infrastructure, and socio-economic development in the region. Spanning a combined area of 222,236 square kilometers, these territories lie between latitudes 32°17′N to 37°05′N and longitudes 72°31′E to 80°20′E (Balasubramanian, 2016; P. Office of Additional Director General of Meteorology (Research), 2014). Topographically, the region is dominated by mountains, with four major Himalayan ranges: Karakoram, Zanskar, Ladakh, and Pir Panjal. The Karakoram range in the north includes K2 (8,611 m), the world’s second-highest peak, and other notable summits like Gasherbrum (8,570 m) and Masharbrum (7,827 m) (Kashmir Travels, 2024). The Zanskar range, part of the Tethys Himalaya, separates the Indus and Kashmir valleys, with elevations reaching up to 6,000 m. The Ladakh range, north of the Indus River, merges with the Kailash range in Tibet and includes peaks above 5,000 m (P. Office of Additional Director General of Meteorology (Research), 2014). To the south, the Pir Panjal range (approx. 4,100 m) forms a climatic barrier between the Jammu and Kashmir regions. Jammu features plains and forested hills merging into the Shivalik and Pir Panjal ranges (Ali et al., 2023). The Kashmir valley, lying between Pir Panjal and Zanskar ranges, is surrounded by snow-capped mountains rising up to 4,900 m, and has an average elevation of 1,615 m (P. Office of Additional Director General of Meteorology (Research), 2014). Ladakh, a cold desert with elevations averaging above 3,650 m, accounts for nearly two-thirds of the region’s total area (P. Office of Additional Director General of Meteorology (Research), 2014). This varied topography results in sharp climatic gradients over short distances, complicating weather prediction. Several perennial rivers, including the Indus, Jhelum, and Chenab, originate in these mountains and sustain the region’s hydrology.

The meteorological profile of the Jammu, Kashmir, and Ladakh regions is marked by significant spatial and temporal variability due to their diverse topography and altitude gradients. Jammu experiences a subtropical climate with hot summers, mild winters, and substantial monsoon rainfall between June and September (Kashmir Travels, 2024). In contrast, the Kashmir Valley features a temperate climate, characterized by cold winters, moderate summers, and two major precipitation periods—winter snowfall from western disturbances (December to March) and moderate rainfall during spring and autumn. Ladakh, situated at a much higher altitude, falls under a cold desert climate with extremely low annual precipitation (less than 100 mm on average) and large diurnal temperature variations (Ali et al., 2023). The region receives minimal monsoonal influence and is instead affected by western disturbances, particularly in the form of snowfall. Seasonal and interannual variability in both temperature and precipitation is pronounced across these regions, presenting significant challenges for reliable forecasting. This meteorological diversity necessitates models capable of capturing complex, localized patterns in both short- and medium-term forecasts.

Accurate weather forecast plays a crucial role in various sectors, including agriculture, tourism, and construction. In addition, will aid famers in crop planning, risk mitigation, and water resource management for optimized yields and resilience. Numerical weather prediction (NWP) is a process of simulating weather conditions, integrating observed data and running simulations to project upcoming weather patterns. This involves solving complex systems of non-linear mathematical equations, derived from specific mathematical models, with the help of powerful supercomputers (Frnda et al., 2022). Meteorological characteristics like temperature and precipitation gathered from a specific location over a period of time yield quantitative data. This quantitative data can be utilized by NWP systems to gain insights into the scientific principles governing atmospheric processes and to make predictions about future atmospheric conditions. However, the performance of NWP is restricted by several factors like extensive resource utilization, inherent uncertainties and limited model resolution. On the other hand, reliability of statistical methods like auto-regressive moving average (ARMA) and its variants depends on the data quality and their ability to effectively depict time-series (TS) data with non-stationary behaviors across seasons (Poornima & Pushpalatha, 2019).

Over the past few decades, deep learning (DL) techniques have proven as a transformative force in advancing the accuracy and efficiency in weather and rainfall forecasting (Waqas et al., 2023). According to Hayati & Mohebi (2007), Waqas et al. (2024) data-driven modeling systems can be employed to decrease the computational requirements of NWP processes. This can be ascribed to their capacity of estimating the intrinsic patterns and dynamics of TS data even without prior knowledge of the parameters, while also accounting for the uncertainty present in observations and system noise. DL techniques such as recurrent neural networks (RNNs) and long short-term memory (LSTM) models are regarded as better suited for handling weather data due to their strong capability to effectively process TS data (Kratzert et al., 2018; Kashiwao et al., 2017). By incorporating input, output, and forget gates, LSTM-RNN achieves a network that can effectively maintain state and propagate gradients over extended time periods in a stable manner.

This study proposes a DL based predictive modeling framework for forecasting T&R that will benefit the local community and stakeholders of J&K and Ladakh by helping them adapt to the unique weather conditions of the area, making their daily life more predictable and safe. The model will also aid the governing bodies of J&K and Ladakh by enhancing foreplanning, risk mitigation, and resource management, ultimately contributing to the economic growth and development of the local population. The proposed model utilizes past values of local temperature and rainfall (T&R) of J&K and Ladakh within a defined time-step as input to forecast the same for the upcoming days. Using weather data of past 23 years from local weather stations will allow the model to capture and model the unique properties exhibited by the weather of J&K and Ladakh.

Related work

Over time, various techniques have been developed to forecast weather conditions, particularly using T&R. A study by Han et al. (2021) trained and tested several neural networks, including feedforward neural networks (FFNN), recurrent neural networks (RNN), long short-term memory (LSTM), and gated recurrent units (GRU), to create a model capable of generating weather predictions reflective of local conditions. Using airport data from the National Solar Radiation Database (NSRDB) and validating it with on-site measurements, GRU model achieved the lowest mean squared error (MSE) of 2.96 among all models. In a separate study aiming to improve weather forecasts for five cities in China, Nketiah et al. (2023) proposed five multivariate time series models specifically designed for atmospheric temperature prediction, based on RNN architectures. The experimental results showed that using the LSTM-RNN model yielded the lowest prediction error for atmospheric temperature compared to baseline models. A new DL model called the spatial feature attention long short-term memory (SFA-LSTM), introduced by Suleman & Shridevi (2022), effectively captured both spatial and temporal relationships among multiple meteorological features for temperature forecasting. Comparative analysis with baseline models demonstrated that the SFA-LSTM achieves state-of-the-art prediction accuracy, offering the additional benefit of enhanced interpretability of spatial features. To capture localized patterns and dependencies contributing to fluctuations in maximum and minimum temperatures, a study by Karevan & Suykens (2020) introduced the transductive-LSTM (T-LSTM) model. This approach, by taking surrounding data points into account, allowed for more accurate temperature forecasts by leveraging the context provided by nearby data points. In another study, Mi (2023) used summer temperature data from Seoul, Korea, spanning from 2013 to 2017, incorporating input parameters such as average temperature, solar radiation, average relative humidity, average wind speed, and average latent heat flux. This study compared the predictive capabilities of linear regression and LSTM models for daily average temperature forecasting, finding that the LSTM model had a lower mean absolute error of 0.19 compared to the linear regression model’s 0.89. To make use of the rapid data processing capabilities of CNNs with the memory function of LSTMs for time series temperature data, Guo (2023) proposed a hybrid CNN-LSTM model. Experimental results showed that this hybrid model outperformed individual CNN and LSTM models, achieving higher temperature prediction accuracy. A similar study by Mung & Phyu (2023) used three DL models—CNN, LSTM, and a CNN-LSTM ensemble—to predict features such as minimum temperature, maximum temperature, humidity, and wind speed. Using root mean square error (RMSE) as the comparison metric, the study found that the ensemble model achieved the lowest RMSE compared to the individual models. Research by Nugraha, Ariawan & Arifin (2023) explored LSTM’s application in weather forecasting using a dataset from the Serang Maritime Meteorological Station, covering variables such as temperature, humidity, sunshine, and wind speed from January 1, 2018, to October 28, 2023. Their LSTM model achieved an RMSE of 0.37 for temperature, 0.72 for wind speed, 2.79 for sunlight, and 5.05 for humidity In another study, Fan et al. (2022) developed a generalized CNN-LSTM model using an encoder-decoder architecture to predict temperature, humidity, rainfall, and wind speed. Trained on the IEEE Big Data IARAI’s Weather4cast 2021 dataset, the model showed a significant reduction in loss as assessed by MSE values. The study by Chen, Huang & Yang (2023) proposed an LSTM model integrating multiple linear regression and Pearson’s correlation coefficients to improve weather prediction for aviation safety. This model, trained on a numerical dataset of 10 weather parameters with key features (sea pressure, dew point temperature, and relative humidity) identified through feature selection, improved its forecasting accuracy from an RMSE of 4.0274 to 2.2215 and a MAPE reduction from 23.0538% to 5.0069%. A comparison study by Gong et al. (2022) evaluated ConvLSTM (LSTM with convolutional filters) against an advanced Stochastic Adversarial Video Prediction (SAVP) network for temperature forecasting. The models were assessed using metrics like MSE, anomaly correlation coefficient (ACC), and structural similarity index (SSIM), revealing that the SAVP model outperformed ConvLSTM with an MSE of 2.3, ACC over 0.85, and SSIM around 0.72. Lastly, study Li et al. (2023) employed LSTM and its deep variant, Deep LSTM (DLSTM), to forecast daily air temperature. Trained on 1,097 weather data points from central and southern regions of Tabriz, Iran, from 2017 to 2019, the DLSTM model achieved an RMSE of 0.08 and an R-squared value of 0.99.

In summary, although prior studies have demonstrated the utility of ML and DL techniques in weather forecasting, they often lack contextual adaptability, overlook complex terrains, or fail to comprehensively evaluate different modelling configurations. This study addresses these limitations by focusing on the climatically diverse Himalayan regions of Jammu, Kashmir, and Ladakh, and by comparing both SISO and MIMO LSTM-based architectures using long-term T&R datasets. In doing so, it offers deeper insights into model suitability and forecasting performance under highly variable geographic and climatic conditions.

Materials and Methods

The model flowchart used in this study is presented in Fig. 1. It includes T & R data collection and preprocessing phase, model configuration (Hochreiter & Schmidhuber, 1997; Werbos, 1988), training and evaluation. The final stage of the experimentation involves assessing and comparing the performance of the trained models.

Models used in the current study

The proposed T&R model leverages advanced DL techniques, specifically RNN and LSTM, to improve prediction accuracy. The following subsections offer a detailed explanation of the RNN and LSTM models.

Recurrent neural network

A traditional recurrent neural network (RNN) (Hochreiter & Schmidhuber, 1997) can be viewed as an enhanced version of a feed-forward NN, designed with an internal memory mechanism to process sequential inputs. RNNs are capable of capturing dependencies from earlier elements in a sequence, allowing them to predict cumulative metrics from sequences of varying lengths, particularly those involving a time-based component (Bukhari & Pandit, 2024). They do this by preserving information in a sequential order i.e., at each time step, the RNN receives the current input along with contextual information passed forward through its feedback connections.

Figure 2 presents a simple RNN architecture, where inputs from time step t to t+1 are denoted as x_t to x_t+1, and the corresponding outputs are represented as y_t to y_t+1. The hidden states, labeled f_t−1 to f_t+1, capture the intermediate information, while the weights for the hidden states, inputs, and outputs are represented by θ, α, and β respectively. RNNs utilize shared parameters across time steps, meaning the same set of weights is applied throughout each layer of the network. Despite being shared, these weights are still refined during training through backpropagation and gradient descent, which helps the model improve its learning performance over time (Bukhari, 2024). A detailed explanation of RNN functionality is provided in Algorithm S1. In RNNs, activation functions are applied solely to the hidden layers. The input layer simply receives and passes the data forward without performing any computation. Without non-linear activation functions, even deep NNs would behave like extended linear regression models, lacking the capability to learn complex patterns from real-world data. To overcome this, commonly used activation functions include ReLU (Rectified Linear Unit), tanh, Sigmoid, and Softmax.

• Backpropagation through time

Backpropagation through time (BPTT) is an advanced form of the traditional backpropagation algorithm, specifically adapted for training RNNs. In BPTT, the RNN is unfolded across its time steps, allowing gradients to be propagated backward through this temporal sequence (Bukhari & Pandit, 2024). This approach enables the model to capture and learn sequential dependencies by adjusting its weights based on the cumulative error over multiple time steps (Hochreiter & Schmidhuber, 1997). A simplified overview of the BPTT process is outlined in Algorithm 2 (provided as Algorithm S2). A simplified breakdown of the steps involved in BPTT is given in Algorithm 2 (provided as Algorithm S2).

Long short-term memory

The limitations of traditional recurrent neural networks (RNNs) such as the vanishing gradient problem, which makes it difficult for the network to capture long-term dependencies, and the exploding gradient problem, which can cause instability during training are mitigated by the use of LSTM networks. LSTM, an advanced variant of RNN, was designed specifically to overcome these challenges, allowing the network to effectively learn from long-term sequential data. LSTM units are structured around a specialized memory cell, which serves as the core of the architecture. This memory cell allows the network to maintain and update information over long time periods, enabling it to better understand the temporal dependencies in the data. The memory cell is updated and modified by three primary gating mechanisms: the input gate, the forget gate, and the output gate as shown in Fig. 3. These gates function collectively to control the flow of information into, out of, and within the memory cell, making LSTM particularly effective at handling time-series data such as T&R sequences. The input gate controls how much of the incoming data at each time step (i.e., the current input and the previous hidden state) should be incorporated into the memory cell. It is a sigmoid function that outputs values between 0 and 1, determining which values will be updated in the cell state. The stronger the activation of the input gate, the more new information will be incorporated. The forget gate, another sigmoid function, determines what portion of the previous cell state should be discarded or retained. Essentially, it tells the network which past information is still relevant to be remembered and which can be forgotten. This is particularly important in time-series forecasting, where certain past events may have a more significant influence on future outcomes than others. By selectively forgetting less useful information, the LSTM can focus on more relevant data points. The output gate governs the flow of information from the memory cell to the output of the LSTM unit. It computes the hidden state, which is a function of the current memory cell and the output from the input gate. The hidden state is then passed to subsequent time steps and contributes to the final prediction. It is also used in the subsequent layers of the network for further processing. This gating mechanism enables the LSTM to learn what information to remember and what to forget across different time scales, making it especially suitable for modelling meteorological time series data, where both short-term variability and long-term seasonal trends are critical. The working of LSTM is explained using algorithm 3 (provided as Algorithm S3).

Results and discussion

Conclusion

This study presents an innovative approach to T&R forecasting, specifically tailored for the Jammu & Kashmir, and Ladakh regions of India, where volatile weather patterns have a significant impact on local livelihoods, socio-economic development, and disaster preparedness. The proposed DL-based model, utilizing RNN and LSTM, demonstrated promising results in accurately predicting short-term variations in T&R. Through the comprehensive analysis of meteorological data obtained from IMD and the application of advanced DL algorithms, this research addresses the unique weather challenges faced by these regions, characterized by extreme and unpredictable weather conditions. By providing more accurate and localized weather forecasts, the proposed model empowers decision-makers, agricultural practitioners, tourism operators, and other stakeholders to better adapt to the region’s weather dynamics. Moreover, the model shall aid governing bodies in enhancing foreplanning, risk mitigation, and resource management, ultimately contributing to the economic growth and development of the local population. The comparative analysis between RNN and LSTM models offers valuable insights into their effectiveness for weather forecasting applications. While both models demonstrate strong performance, particularly in the SISO configuration, the RNN model consistently outperforms the LSTM model in capturing the temporal relationships of weather data. This finding highlights the importance of selecting appropriate model architectures based on the specific characteristics of the data and the forecasting objectives. Furthermore, the validation of the proposed model against existing benchmark models displays its superiority in terms of accuracy and performance. The proposed RNN model achieves significantly lower MAE, RMSE and MSE values compared to previous studies, indicating its efficacy in capturing the intricate patterns of T&R variations. The applicability of this research extends beyond academic inquiry, offering practical benefits to various sectors, including agriculture, tourism, construction, and disaster management. Accurate T&R forecasting enables stakeholders to make informed decisions, optimize resource allocation, and mitigate risks associated with adverse weather conditions. By enhancing the resilience of communities and supporting sustainable development initiatives, the proposed models shall contribute to the overall well-being and prosperity of the J&K and Ladakh regions. Future research directions may include further refinement of the proposed model, integration of additional meteorological variables, and exploration of ensemble forecasting approaches for enhanced predictive accuracy and reliability.

While the study focused on continuous rainfall prediction using regression-based metrics, assessing model performance from a classification perspective such as distinguishing rain from no-rain events or evaluating accuracy across rainfall intensity categories remains equally important, particularly for operational applications in water resource management and disaster preparedness. Incorporating such event-based metrics, including probability of detection (POD) and critical success index (CSI), is an important direction for extending this research and enhancing the practical utility of DL models in hydrometeorological forecasting.

Despite the encouraging performance of the proposed models across various stations, several limitations warrant consideration. This study focused exclusively on historical univariate time series data, without incorporating spatial or contextual meteorological inputs, which may limit the generalizability of the findings across broader climatic regions. Moreover, the effectiveness of DL models is inherently influenced by the volume and consistency of data; stations with comparatively sparse records may have constrained model learning. Although both SISO and MIMO forecasting configurations were evaluated, their real-world application would benefit from further validation under more diverse temporal and spatial scenarios. Addressing these limitations through the inclusion of spatially distributed features, richer datasets, and exploring hybrid architectures such as CNN-LSTM present a promising direction for future research.

Supplemental Information