FIGURE 6 - available via license: Creative Commons Attribution 4.0 International
Content may be subject to copyright.
Source publication
High-resolution temperature forecasting can often prove to be challenging for conventional machine learning models as temperature is highly seasonal and varies with the time of the year as well as with passing hours of the day. In most cases, only the daily extremes or mean temperatures are provided by temperature forecasting methods. However, with...
Context in source publication
Similar publications
Federated learning is a distributed machine learning framework that enables a large number of devices to cooperatively train a model without data sharing. However, because federated learning trains a model using non-independent and identically distributed (non-IID) data stored at local devices, the weight divergence causes a performance loss. This...
Citations
... However, with the increased availability of high-frequency data and enhanced computing power, mainly through graphics processing units, there is now the potential to utilize hourly input data for daily temperature forecasting, surpassing the previous constraint of exclusively employing daily input data. Both hourly and daily patterns can contribute to forecasting daily temperatures (Haque et al. 2021). Given the abundance and detail of these data, efficient and effective processing becomes crucial. ...
... Subsequently, in the last layer, the dense layer, flattened data from the pooling stage is transformed into a 1D output sequence. A notable advantage of 1D CNNs is their potential for low-cost hardware implementation, primarily involving 1D convolutions, which consist of additions and scalar multiplications (Haque et al. 2021). ...
Air temperature is a crucial climatic indicator that significantly impacts various sectors, including the environment, hydrology, agriculture, and disaster management. Accurate and timely air temperature forecasting is essential for effective risk management and future planning. This study investigates the performance of Deep Learning (DL) models for nowcasting air temperature in various regions of Pakistan. We utilize hourly temperature data (2018–2023) from four meteorological sites (Murree, Swat, Multan, and Sukkur), representing different climate conditions. The models compared in this study include Long Short-Term Memory (LSTM), Convolutional Neural Networks (CNN), Feed Forward Neural Networks (FNN), and a hybrid CNN-LSTM architecture. The performance of these models is evaluated using several statistical criteria (MSE, MAE, RMSE, and MAPE) and visual comparisons. The results indicate that while the LSTM model performs best, the CNN, FNN, and hybrid CNN-LSTM models also show considerable promise. However, it can be concluded that the LSTM model outperforms other models. These findings highlight the adaptability of DL algorithms in predicting temperature across various climatic scenarios. The implications of this study are significant for sectors such as agriculture, transportation, and disaster relief, which depend on accurate temperature forecasts for effective resource allocation and climate risk management. By advancing nowcasting technologies in Pakistan, this research contributes to enhancing resilience to weather-related challenges.
... To address this limitation, Fernandes et al. [18] introduced combined LSTM networks for traffic flow forecasting, demonstrating that LSTMs can effectively predict traffic flow for multiple future time steps by addressing key model aspects such as input features and time frames. Haque et al. [19] demonstrated that hybrid models like GRU-LSTM outperform single-layer models in high-resolution temperature forecasting by capturing both short-term and long-term trends, with GRU proving consistently robust across diverse locations. Similarly, Hossain et al. [20] combined convolution neural networks (CNN), GRU, and fully connected neural networks for wind energy generation forecasting, demonstrating that hybrid architectures can effectively capture both short-term fluctuations and long-term trends. ...
Sand and dust storms significantly challenge microwave and millimeter-wave communications, particularly in arid and semi-arid regions. Various models have been developed to predict attenuation caused by these storms theoretically and empirically based on two meteorological parameters, namely visibility and humidity. However, these models are found unable to predict most of the attenuation measurements. This study presents a hybrid Machine Learning (ML) model that predicts dust storm attenuation for 22 GHz terrestrial links using meteorological data. The received signal levels were measured for a 22 GHz link over a month in Khartoum, Sudan. The visibility, humidity, atmospheric pressure, temperature and wind speed were also monitored simultaneously by Automatic Weather Station (AWS). The proposed model incorporates XGBoost for feature selection and combines Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) layers to capture both short-term and long-term dependencies in meteorological data. The results demonstrate a strong correlation between meteorological parameters and dust storm attenuation. The model’s performance is validated against the measured data at 22 GHz, outperforming existing empirical and theoretical models. The RMSE for the proposed model is 0.07, while all existing theoretical and empirical models are higher than 0.25. Furthermore, the proposed model demonstrates significant enhancements over the available ML model for dust attenuation prediction. This hybrid ML approach offers a more accurate and robust solution for predicting microwave and millimetre wave attenuation during dust storms, enhancing the reliability of communication systems in affected regions.
... This architecture was optimized by tuning hyperparameters such as learning rate, batch size, and the number of hidden units to minimize prediction errors. The GRU-LSTM model has demonstrated higher accuracy than individual GRU and LSTM models and lower RMSE values than other prediction methods [17], [18]. ...
... The GRU-LSTM Hybrid model was created on the Google Collab platform using the Python language and various Python libraries, such as time, numpy, pandas, sklearn, TensorFlow, and Keras. Fig. 5 shows the architecture of the GRU-LSTM Hybrid model [17]. ...
... For example, Support Vector Machines (SVMs) are often chosen for their ability to handle highdimensional data and effectively model nonlinear relationships [42][43][44]. Meanwhile, deep learning architectures-such as Recurrent Neural Networks (RNNs) [45][46][47] or Long Short-Term Memory (LSTM) networks [48][49][50]-can incorporate temporal dependencies, thereby capturing the complex, time-varying dynamics that characterize T2M-LST interactions. Moreover, these models can be enhanced by integrating land cover information (e.g., NDVI or NDBI), which helps account for the influence of vegetation and urbanization on near-surface temperature patterns [51][52][53][54][55][56][57]. ...
This study develops and compares three deep learning methods—LSTM, TCN, and N-BEATS—for estimating near-surface air temperature (T2M) from satellite-derived land surface temperature (LST) and land cover metrics such as NDVI and NDBI. By incorporating temporal context through varying look-back windows, these models substantially outperform non-temporal baselines, reducing root-mean-square error (RMSE) from around 2.6–2.8°C to below 1.8°C, and underscoring the value of historical LST observations for capturing the evolving surface–air temperature relationship. Longer lags generally improve accuracy, although N-BEATS performance plateaus beyond a certain window, reflecting both diminishing returns and practical limitations linked to missing cloud-free satellite data. Seasonal and diurnal evaluations show higher errors in spring and midday hours, likely due to rapid vegetation changes and stronger physical and dynamical processes that make T2M less predictable. Spatially, stations with denser vegetation exhibit elevated errors, suggesting that transpiration and canopy effects complicate the LST–T2M linkage. For extreme-event detection, LSTM provides the fewest false alarms (highest precision), N-BEATS captures the most extremes (highest recall), and TCN offers the best overall balance in precision and recall (highest F1). While cloud-free satellite coverage remains a limitation, future work could explore adaptive lag strategies, additional data sources, and more advanced data-fusion techniques. These results highlight that satellite-based temperature monitoring, when combined with suitable deep learning architectures, can reliably estimate T2M based on LST, further addressing gaps in near-surface observations and facilitating the detection of critical T2M extremes. This framework has direct applications in heat-warning systems, resource management, precision agriculture, and urban climate adaptation, and stands to benefit further from ongoing advancements in satellite sensing technology.
... Predictions with high resolution and more data will help in planning and policy. Temperature prediction is an important feature that increases the prediction horizon for various life applications [29]. ...
Climate change is a global challenge that requires serious attention from various parties, including the government. The existence of surface temperature and various other parameters is certainly closely related to climate change. In this context, this study was conducted to identify the best model in predicting urban land surface temperature in the Jakarta area, as one of the steps to understand and deal with the impacts of climate change. The research data used comes from MERRA-2, NASA, which provides datasets for various climate analyses. A comparison of ARIMA, SVR, LSTM, and ANN methods was conducted to evaluate the performance of each model in forecasting land surface temperature. The results show that the Long-Short Term Memory (LSTM) model provides the best performance with MAPE and values of 0.8381 and 0.8628. This model has an advantage over other models because it can remember various information that has been stored for a long period of time and can delete irrelevant information. This shows that LSTM is effective in capturing the pattern and variability of the Earth's surface temperature in the Jakarta area. Based on these findings, the government is expected to take concrete steps to address the impacts of climate change, especially issues related to increasing urban land temperature in Jakarta, such as reducing the use of private vehicles and switching to public transportation, expanding green open space, and relocating residents to reduce density.
... The proposed approach generates precise predictions with low RMSE and MAE values of 0.0813 and 0.1359, respectively. The GRU model combines LSTM to forecast hourly temperatures based only on past temperature data [19]. The findings indicated that the GRU-LSTM hybrid model had the lowest RMSE of 1.691°C, followed by CNN-LSTM with an RMSE of 1.7485°C. ...
... LSTM has an advantage over standard RNN as it can retain relevant information for extended periods [9]. This feature addresses the vanishing and exploding gradients that RNNs encounter while handling lengthy sequences [19]. An LSTM unit comprises a memory cell and three gates: the input gate, the output gate, and the forget gate, as depicted in Fig. 1. ...
... Time series prediction has a wide scope and challenges in research for so many years, with significant applications including disease spread prediction (Abbasimehr and Paki, 2020;Chandra et al., 2022), weather conditions (Haque et al., 2021;Dubey et al., 2024), air pollutants (Drewil and Al-Bahadili, 2022), stock price prediction (Kumar et al., 2022;Zhou et al., 2022) and many more. Ample research work has been committed in recent decades to improve and develop ML Models for prediction. ...
... The study demonstrates that Cauchy-Exploration strategy BAS(CESBAS) & ANFIS achieve better COVID-19 time series prediction than other optimizations such as GA and PSO (Zivkovic et al., 2021). The benefit of ensemble RNN models has been applied (Haque et al., 2021) in Beijing Temperature prediction. It was demonstrated that the CNN-LSTM parallel Network obtains the lowest RMSE. ...
Currently, Deep Learning (DL) with the Recurrent Neural Networks (RNN) variants is being applied successfully in many domains of Engineering for prediction. In view of the demand for precise forecasting and the aid of Artificial Intelligence Tools, time series prediction reveals a vital task in decision-making and risk assessment. However, the application of novel Recurrent DL models for obtaining an accurate prediction of time series is yet to be explored. Recent trends reveal that Hybrid Neural Networks and DL models are appropriate for time series forecasts. At the same time, the model's selection and the hyperparameter's tuning can greatly impact its performance. To address this problem, a parallel long-term memory (PLSTM) model integrated with Bayesian hyperparameter optimization (PLSTM-BO) is proposed for time series prediction. The model is tuned in terms of key parameters, including the number of neurons, dropout, learning rate, and optimization technique. The model's performance is assessed using the SARS-COVID-19 cumulative cases, deaths, recovery cases, and NIFTY 50 stock closing price time series dataset. The obtained results convey that the current model exhibits remarkable performance compared to existing models.
... Dans un autre contexte, les réseaux neuronaux artificiels (Artificial Neural Network ou ANN) ont été exploités pour la prédiction de la température horaire de l'air (Li et al. 2020;Haque et al. 2021; Gong et al. 2022). Et il ressort de ces études qu'à l'échelle régionale, les modèles d'apprentissage profond apportent des prévisions bien plus précises que celles de l'apprentissage machine (Abubakar et al. 2016). ...
[FR] Actuellement, le Bassin du Congo représente le centre le plus important en termes de concentration en biodiversité, surtout
avec la déforestation croissante observée en Amazonie. Les modèles climatiques disponibles sont majoritairement à des échelles plus
grandes et peu d’entre eux se concentrent sur des zones spécifiques du Bassin du Congo, comme la localité de Makokou au Gabon. Une
nouvelle approche est donc nécessaire pour prédire les changements de température dans cette région particulière. Bien qu’il existe
quelques travaux portant sur la prédiction des températures, la majorité n’utilisent pas les algorithmes d’apprentissage profond. Cette
contribution vise à comparer les prédictions d’un modèle de mémoire à court et long terme (LSTM) avec celles issues de la combinaison
Transformée en Ondelettes et LSTM (TO-LSTM). Le modèle LSTM développé comprend deux couches LSTM, deux couches Dropout (à
un taux de 50%) et une couche Dense pour afficher la valeur prédite. Le modèle TO-LSTM présente des résultats supérieurs à ceux du
modèle LSTM, avec une racine carrée d’erreur quadratique moyenne de 0,45 °C, une erreur absolue moyenne de 0,35 °C et un coefficient
de corrélation de Spearman de 0,97 °C. Ces résultats soulignent l’importance d’utiliser des approches avancées pour améliorer les
prévisions climatiques dans des zones cruciales pour la conservation de la biodiversité. La précision accrue des prévisions pourrait aider à mieux anticiper et atténuer les impacts des changements climatiques locaux, contribuant ainsi à la gestion durable de cette région
écologiquement sensible.
MOTS-CLEFS: changements climatiques, apprentissage profond, modèle de mémoire à court et long terme LSTM, Transformée en
Ondelettes, prédiction des températures, biodiversité, Bassin du Congo.
[ENG] Currently, the Congo Basin represents the most important center in terms of biodiversity concentration, especially with the
increasing deforestation observed in the Amazon. The available climate models are mostly at larger scales, and few of them focus on
specific areas of the Congo Basin,such as the locality of Makokou in Gabon.A new approach is therefore needed to predict temperatures
changes in this particular region. Although some work focus on temperature prediction, most do not use deep learning algorithms. This
contribution aims to compare the predictions of a Long Short-Term Memory (LSTM)model with those from the combination of Wavelet
Transform and LSTM (WT-LSTM). The developed LSTM model includes two LSTM layers, two Dropout layers(with a rate of 50 %) and a
Dense layer to outpout the predicted value. The WT-LSTM model shows superior results compared to the LSTM model, with a root mean
square error of 0.45 °C, a mean absolute error of 0.35 °C, and a Spearman correlation coefficient of 0.97 °C. These results highlight the
importance of using advanced approaches to improve climate forecasts in areas crucial for biodiversity conservation. The increased
accuracy of predictions could help better anticipate and mitigate the impacts of of local climate change, thereby contributing to the
sustainable management of this ecologically sensitive region.
KEYWORDS: climate change, deep learning, Long Short-Term Memory model (LSTM), Wavelet Transform, temperature prediction,
biodiversity, Congo Basin
... Many studies have applied LSTM and GRU in the context of temperature forecasting [36]- [47]. However, the choice between these two models often depends on factors such as dataset characteristics, model complexity, and computational efficiency. ...
... However, LSTM has some drawbacks, [36]. LSTM also has longer training times and demands more computational resources compared to GRU algorithms [47]. On the other hand, GRU has advantages in training speed and memory efficiency, making it faster and lighter than LSTM [42], [45]. ...
... 4, No. 3, 2024 the model to selectively update and forget information, thus capturing relevant patterns in the input data [44]. In temperature prediction cases, GRU also handles long-term dependencies effectively, useful in forecasting temperatures where seasonal patterns and long-term trends influence the data [44], [47]. However, GRU may not be as accurate as LSTM for very long sequence data and is less flexible in handling extremely long memory spans [45]. ...
Temperature forecasting is a crucial aspect of meteorology and climate change studies, but challenges arise due to the complexity of time series data involving seasonal patterns and long-term trends. Traditional methods often fall short in handling this variability, necessitating more advanced solutions to enhance prediction accuracy. LSTM and GRU models have emerged as promising alternatives for modeling temperature data. This study is a literature review comparing the effectiveness of LSTM and GRU based on previous research in temperature forecasting. The goal of this review is to evaluate the performance of both models using various evaluation metrics such as MSE, RMSE, and MAE to identify gaps in previous research and suggest improvements for future studies. The method involves a comprehensive analysis of previous studies using LSTM and GRU for temperature forecasting. Assessment is based on RMSE values and other metrics to compare the accuracy and consistency of both models across different conditions and temperature datasets. The analysis results show that LSTM has an RMSE range of 0.37 to 2.28. While LSTM demonstrates good performance in handling long-term dependencies, GRU provides more stable and accurate performance with an RMSE range of 0.03 to 2.00. This review underscores the importance of selecting the appropriate model based on data characteristics to improve the reliability of temperature forecasting.
... CNNs are excellent for capturing spatial (two-or threedimensional) patterns, while LSTMs are excellent for sequences or temporal dynamics. Haque, Tabassum and Hossain (2021) conducted a comparative study of six different deep network architectures: gated recurrent unit (GRU),convolutional neural network (CNN), simple recurrent neural network (SRN), long-short term memory (LSTM), and two hybrid models. The paper reported a RMSE value of 1.1729, 1.2010, 0.9764, 1.0270, 0.8575 and 1.0277 for SRN, GRU, LSTM, CNN, CNN-LSTM and GRU-LSTM respectively for 1-hour ahead prediction which shows the excellent temporal and spatial extraction capabilities of the hybrid CNN-LSTM approach. ...
Accuracy and model compactness are essential requirements for weather forecasting models designed for operation on low-power embedded devices. This study developed Mixed-Input Residual Network (MIRNet), a compact temperature-forecasting deep neural network model. MIRNet integrates stacked bidirectional long short-term memory layers using concatenated 1-dimensional and 2-dimensional convolutional layers to improve model accuracy. MIRNet was trained and tested on two datasets: one, IfeData, comprising historical weather data from Ile-Ife, Nigeria and the other a standard weather forecasting dataset called the Jena dataset. Training was carried out using 100 epochs of data partitioned in the standard 80:20 ratio, with an adaptable learning rate strategy. The model was tested for Nth-hour-ahead prediction for 1N24; where N are natural numbers, and performance quantified using metrics such as mean absolute percentage error (MAPE) and mean square error (MSE). The model was also implemented on a Raspberry Pi 4 device with a 1.8 GHz 64-bit quad-core ARM Cortex-A72 processor. The model achieved a MSE of 1.00 x 10-3 on the IfeData dataset, and 1.23x10-4 on the Jena dataset for 1-hour ahead forecasting. This is currently the best verifiable result achieved on the Jena dataset by any prediction model globally. For Nth hour ahead forecasting, MIRNet achieved an MSE generally below 2.0x10-3 for all values of N on the standard Jena dataset. The MSE of MIRNet for N-sequential 1-hour ahead and single Nth hour predictions using the Jena dataset reveal quadratic and linear relationships with N respectively. The model compares favourably with existing models for multi-hour predictions. The developed model is compact and has good forecasting properties.
