Comparing Deep Learning Approaches for Weather Forecasting: Insights from the PRECEDE Project Maira Aracne1,* , Tommaso Ruga2,* , Camilla Lops3,* , Deborah Federico4 , Luciano Caroprese3 , Ester Zumpano2 , Sergio Montelpare3 , Mariano Pierantozzi3 , Francesco Dattola4 , Pasquale Iaquinta4 , Miriam Iusi4 , Raffaele Greco4 , Marco Talerico4 , Valentina Coscarella4 , Luca Legato4 , Ivana Pellegrino4 , Sonia Bergamaschi5 , Mirko Orsini5 , Riccardo Martoglia5 , Andrea Livaldi5 , Abeer Jelali5 and Simone Sbreglia5 1 University Leonardo da Vinci, Piazza San Rocco, 2, Torrevecchia Teatina, Italy 2 DIMES Department, University of Calabria, Via Ponte Pietro Bucci, Rende (CS), Italy 3 Engineering and Geology Department, University G. d’Annunzio of Chieti-Pescara, Viale Pindaro 42, Pescara, Italy 4 e way Enterprise Business Solutions, Via Francesco de Francesco 19, Cosenza, Italy 5 DataRiver Srl, Via Emilia Est, 985, Modena, Italy Abstract Accurately predicting weather conditions in advance is crucial across various sectors. It informs decision-making in agriculture, enables preparation for potential natural disasters, optimizes renewable energy management, and helps reduce energy waste and inefficiencies. In the current historical context, achieving maximum forecast precision has become more critical than ever. Artificial intelligence is transforming the methods and tools we use to reach this goal, paving the way for unprecedented advancements. Its main advantage lies in the ability to manage and evaluate huge amounts of data identifying complex patterns and correlations that could escape from human analysis. The system’s fast processing capabilities constitute another fundamental aspect, producing territory-specific forecasts that consider both local micro-climates and distinctive geographical features. The present work aims to compare different deep learning approaches applied to weather forecasting, conducted as part of the PRECEDE Project. Recurrent neural networks are analyzed, in particular Gated Recurrent Units, and Temporal Convolutional Networks, known to be two architectures specialized in modelling data sequences over time horizons. The study highlights the performance of neural networks in enhancing the outputs of the MM5 weather model, a regional mesoscale model, over one-, two-, and three-day time horizons. Furthermore, the work explores the strengths and limitations of each approach, providing insights into their effectiveness, which serve as a foundation for guiding future research and practical applications of deep learning in weather forecasting. Keywords Weather forecasting, Deep Learning, Machine Learning, Mesoscale Model 5, Renewable energy prediction 1. Introduction terns. Deep Learning has improved accuracy by analyzing large datasets and identifying complex temporal patterns. Weather forecasting plays a crucial role in decision-making These advances help optimize energy systems while main- across various sectors, including agriculture, disaster man- taining grid stability and meeting prosumer needs. agement, energy optimization and urban planning. Reliable DL methods, including Recurrent Neural Networks predictions reduce risks, enhance efficiency and support (RNNs), Gated Recurrent Units (GRUs), and Temporal Con- the transition to sustainable energy systems through bet- volutional Networks (TCNs), excel at modeling temporal ter management of renewable sources. As global energy data sequences and capturing climate data trends. Addition- demand rises, advanced forecasting methods help optimize ally, traditional models like the Fifth-Generation NCAR- wind and solar energy usage within decentralized models /Penn State Mesoscale Model (MM5) maintain their im- like energy communities - local networks where individuals portance due to operational dependability, offering solu- both produce and consume energy, essential for reducing tions across multiple sectors. Despite these advancements, fossil fuel dependency. two significant challenges persist in the pursuit of precise Weather forecasting is challenging due to dynamic pat- weather forecasting and its integration with energy systems. The first challenge lies in developing a robust platform ca- Published in the Proceedings of the Workshops of the EDBT/ICDT 2025 pable of efficiently managing, integrating, and analysing Joint Conference (March 25-28, 2025), Barcelona, Spain vast amounts of heterogeneous data from diverse sources. * Corresponding author. This is essential given the intricate relationship between $ m.aracne@unidav.it (M. Aracne); tommaso.ruga@dimes.unical.it (T. Ruga); camilla.lops@unich.it (C. Lops); climatic variables and renewable energy production. The dfederico@eway-solutions.it (D. Federico); luciano.caroprese@unich.it second challenge focuses on leveraging this integrated data (L. Caroprese); e.zumpano@dimes.unical.it (E. Zumpano); to design advanced models and services that can accurately sergio.montelpare@unich.it (S. Montelpare); predict energy production and meet evolving management mariano.pierantozzi@unich.it (M. Pierantozzi); requirements, ensuring both reliability and sustainability. fdattola@eway-solutions.it (F. Dattola); piaquinta@eway-solutions.it (P. Iaquinta); miusi@eway-solutions.it (M. Iusi); In this context, the present work aims to address the lim- rgreco@eway-solutions.it (R. Greco); mtalerico@eway-solutions.it itations of current Regional Climate Models (RCMs), includ- (M. Talerico); vcoscarella@eway-solutions.it (V. Coscarella); ing MM5’s occasional inaccuracies, through a comparative llegato@eway-solutions.it (L. Legato); ipellegrino@eway-solutions.it analysis of deep learning methodologies applied to weather (I. Pellegrino); sonia.bergamaschi@unimore.it (S. Bergamaschi); forecasting. Insights are drawn from the PRECEDE project, mirko.orsini@datariver.it (M. Orsini); riccardo.martoglia@datariver.it (R. Martoglia); andrea.livaldi@datariver.it (A. Livaldi); which explores innovative approaches to enhance predic- abeer.jelali@gmail.com (A. Jelali); simo.sbreglia@gmail.com tion accuracy and optimize energy-related applications by (S. Sbreglia) integrating the strengths of DL models with traditional fore- © 2025 Copyright © 2025 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). CEUR ceur-ws.org Workshop ISSN 1613-0073 Proceedings casting frameworks. Additional details about the project casting advantages that depend on its specific application can be found in [1]. scenario. In detail, machine and deep learning models are imple- For weather forecasting applications, studies in [12] and mented to combine real-time measurements with RCM out- [13] demonstrated TCN models’ effectiveness in predicting puts. The models analyze climate data streams to iden- Global Horizontal Irradiation (GHI) and ten weather param- tify seasonal patterns and short-term trends, improving the eters, respectively. However, while these works addressed prediction of key weather parameters - including solar ra- forecasting horizons ranging from minutes (5, 10, 15, and diation, temperature, relative humidity, and atmospheric 20) to several hours (up to 9), they operate on different time pressure - which in turn leads to more accurate forecasts scales than the here presented research. of renewable energy generation. This comprehensive ap- Despite notable advances in energy management and proach optimizes energy storage and distribution for energy forecasting research, most approaches remain fragmented communities, while improving the efficiency and sustain- rather than converging into comprehensive solutions. While ability of renewable energy systems at both individual and machine learning has shown promising results in weather community scales. prediction [2, 3, 4, 5], solar radiation estimation [6, 14, 15, 16] The paper is structured as follows: after this brief in- and temperature forecasting [7, 17, 18], these advances have troduction, Section 2 reviews the state-of-the-art artificial not been fully integrated into comprehensive energy man- intelligence techniques currently used for predicting cli- agement systems. Furthermore, existing energy optimiza- matic parameters. Section 3 provides an overview of the tion strategies [19, 20, 21, 22] tend to focus on specific com- MM5 model, the weather datasets and the chosen artificial ponents, such as battery management or demand response, neural networks utilized in this study. Finally, the results without addressing the complex, interconnected nature of are presented in Section 4, followed by a discussion of the energy communities. To overcome these limitations, the main findings in Section 5. PRECEDE project introduces an integrated framework that leverages multiple AI techniques through a modular archi- tecture, comprehensively addressing the energy manage- 2. Related Works ment pipeline from data integration to community-scale optimization. Unlike previous approaches limited to spe- Predicting climate variables is notoriously difficult due to cific community data, PRECEDE’s architecture transcends their dynamic nature, thus considerable effort has been these limitations offering a generalizable framework that made to apply Artificial Intelligence (AI) to this challenge. adapts to diverse settings and environmental conditions Consequently, a new field, Deep Learning for Weather Pre- while bridging the gap between climate forecasting and diction (DLWP), has born and it has demonstrated impres- energy optimization. sive results, as shown in [2, 3]. The ability of Neural Net- works to learn complex nonlinear relationships and to pro- cess vast amounts of data simultaneously enables their appli- 3. Background cation in different fields, such as in solar radiation prediction (at both daily and hourly scales), short-term and long-term This section explores both the meteorological and AI aspects wind resource estimation, and in the forecasting of various of our datasets and proposed models. meteorological parameters such as temperature, precipita- tion, cyclones, and humidity [2, 4, 5, 6, 7] 3.1. The Fifth Mesoscale Model (MM5) Among the available architectures, we chose the GRUs and the TCNs due to their specialization in modeling data The Fifth-Generation Penn State/NCAR Mesoscale Model sequences. LSTM and GRU are the two main RNN variants (MM5) is a widely used numerical weather prediction sys- that handle long sequences better than vanilla RNNs. After tem, developed collaboratively by the Penn State Univer- conducting a comparative evaluation between GRU and sity and the National Center for Atmospheric Research LSTM on a sample dataset, our analysis revealed comparable (NCAR). It is designed to simulate mesoscale and regional performance between the two models. GRU has been chosen atmospheric phenomena for both research and operational for its simpler design, faster training, and more efficient forecasting. memory use. MM5 is highly adaptable, offering configurable grid res- The TCNs are well known to outperform RNNs across a olutions, physical parameterizations, and boundary con- broad range of sequence modeling tasks [8]. However, in ditions to suit various meteorological applications. It em- [9] , the GRUs show better prediction capability than TCNs, ploys a 𝜎-coordinate system based on hydrostatic pressure but the problem of the correct tuning hyperparameters is and finite-difference numerical schemes, specifically the opened. Also, the study raises the possibility that results Arakawa-Lamb B-staggering technique, enabling detailed may differ if the lengths of the input or output changes. simulations of convection, radiation, cloud microphysics, The full potential of GRUs and TCNs in climate variable and surface-atmosphere interactions. prediction remains to be explored, as their application in The non-hydrostatic model relies on conservation equa- this domain is still an emerging area of research. tions for momentum and energy, incorporating a tendency Different solutions exist in the field of energy, or power, equation for perturbation pressure. MM5 predicts meteo- forecasting. Shaikh et al. [10] demonstrated that TCNs rological variables like temperature, pressure, wind, solar typically outperform LSTM models. On the other hand, radiation, and cloud cover, making it suitable for diverse the review conducted in [11] studies the possible advan- applications, including short-term weather forecasting, cli- tages and disadvantages of different neural networks for mate studies, air quality management, water resource plan- Photovoltaic (PV) power prediction and it finds that the Mul- ning, and severe weather analysis. tilayer Perceptron, RNNs, Convolutional Neural Network, With portability across computational platforms and ex- and Graph Neural Network architectures have different fore- tensive documentation, MM5 is accessible to users of vary- ing expertise. Additional details about the MM5 system and its key features are available in Grell et al. [15] and Dudhia et al. [16]. 3.2. Datasets Both case studies, Casaccia and Ottana, utilize datasets with measurements collected at 10-minute intervals. The Casac- cia database spans three years, beginning January 1𝑠𝑡 , 2018 at 1:10 am, and includes four physical quantities: GHI, tem- perature, atmospheric pressure, and relative humidity. The Ottana dataset covers one year, starting from May 31𝑠𝑡 , 2021 Figure 1: GRU basic building block. at 17:10, and comprises three physical quantities: GHI, at- mospheric pressure, and relative humidity. The difference in dataset duration arises from their availability and reliability from weather stations in each location. Acquiring contin- uous and high-quality meteorological records remains a challenge, and the selected databases represent the most comprehensive and accurate data accessible for each site. Additionally, the study’s methodological approach accounts for these variations by focusing on relative trends and pat- terns rather than absolute comparisons, thus maintaining the validity of the performance evaluation across locations. For both locations, each measured quantity is paired with its corresponding MM5 system prediction, and all values have been normalized. Distinct datasets were created for each combination of prediction horizon 𝜏 (144 for 1 day, 288 for 2 days, and 432 for 3 days) and target variable i (0 for GHI, 1 Figure 2: TCN basic building block. for temperature, only for Casaccia, 2 for atmospheric pres- sure and 3 for relative humidity). Each of these datasets has been divided into three subsets training set (60%), validation set (20%) and test set (20%). in a single dimension (time series). In this case, the con- volution is made causal, meaning the network learns to 3.3. Artificial Intelligence Techniques predict the output at time 𝑡 by only considering data up to time 𝑡, avoiding the use of future information to predict the 3.3.1. Gated Recurrent Unit present. A key requirement for a forecasting model is that The GRU is a specialized recurrent neural network architec- each output element should depend on all historically pre- ture that excels in modeling sequential data and temporal ceding input elements: the TCN adopts dilated convolutions dependencies. At its core, the GRU cell processes sequential to expand its receptive field without dramatically increasing information through a sophisticated gating mechanism that the number of parameters. This technique ensures coverage selectively retains or discards information at each time step. of extensive sequence portions without losing resolution This mechanism consists of two primary components: the or computational efficiency, even with a small convolution update gate and the reset gate. The update gate balances kernel. Importantly, expanding the receptive field enhances the integration of new information with historical context, the network’s ability to capture long-term dependencies. determining how much of the previous hidden state should To address the vanishing gradient problem, particularly in persist. Meanwhile, the reset gate controls the forgetting deep networks, residual connections are employed, creat- mechanism, allowing the model to discard irrelevant past ing a direct path between the network’s input and output. information. This dual-gate architecture allows GRUs to Currently, TCNs are considered an alternative to RNNs, in- effectively model long-term dependencies while address- cluding their more sophisticated versions, LSTM and GRU. ing the vanishing gradient problem inherent in traditional Some advantages of TCNs are their ability to parallelize RNNs, as both gates work together to precisely control the work, which streamlines the training process, and the ab- temporal evolution of the hidden state. By dynamically sence of a recursive structure, which leads to more stable managing information flow, GRUs maintain an adaptive gradient propagation. memory that evolves with the input sequence, making them particularly effective for tasks like time series forecasting 3.3.3. Deep Reinforcement Learning and natural language processing. To manage energy communities, the global PRECEDE frame- work has adopted a DRL approach, widely considered one 3.3.2. Temporal Convolutional Networks of the most effective methodologies in this field. This ad- The TCN is a neural network specialized in modeling data vanced computational paradigm combines DL architectures sequences, drawing inspiration from the operational mecha- with the Reinforcement Learning (RL) framework, enabling nism of Convolutional Neural Network, which uses filters to robust solutions for complex decision-making processes. recognize patterns in data. However, instead of performing However, it is important to note that the DRL approach was two-dimensional convolution (as with images), it operates not utilized in the analyses conducted here. RL operates on the principle of sequential decision- making, where an agent interacts with an environment First Layer: through an iterative process of observation, action, and re- Data Integration Layer ward. The environment is typically modeled as a Markov (MOMIS Platform) Decision Process (MDP), characterized by a state space 𝑆, an Integration of heterogeneous data sources action space 𝐴, and a reward function 𝑅. At each time step Integrated data 𝑡, the agent observes the current state 𝑠𝑡 ∈ 𝑆 and selects an action 𝑎𝑡 ∈ 𝐴 based on its policy 𝜋(𝑎|𝑠), which maps states Second Layer: to action probabilities. Following the action, the environ- Climate Variables Broadcasting (Deep Learning: GRU, TCN) ment transitions to a new state 𝑠𝑡+1 and provides a reward Accurate climate variable forecasting signal 𝑟𝑡 . The agent’s objective is to learn an optimal pol- icy 𝜋 * that maximizes the∑︀expected cumulative discounted Climate forecasts reward, expressed as 𝐸[ 𝛾 𝑡 𝑟𝑡 ], where 𝛾 ∈ [0, 1] is the Third Layer: discount factor balancing immediate and future rewards. Energy Production Forecasting DRL represents a sophisticated integration of deep learn- (Physical Models) ing architectures with Reinforcement Learning principles, PV energy production estimation designed to handle complex decision-making tasks in high- Production Estimation dimensional spaces. The integration of deep learning en- Fourth Layer: hances the agent’s capability to identify complex patterns Energy Flow Optimization and hierarchical representations, leading to more sophisti- (Deep Reinforcement Learning) cated decision-making strategies. Multi-agent optimization system 4. Discussion Figure 3: PRECEDE system layers overview. To fully understand the PRECEDE project, a brief introduc- tion strictly connected to the layers which compose the 4.1. Experimental Settings proposed system is resumed in Fig. 3. The PRECEDE architecture is composed of four different lay- The model was developed in Python (version 3.11.11 ) using ers, which belong to two different processing steps: the data PyTorch (version 2.5.1+cu121) and a Tesla T4 GPU with 51 integration and processing step and the energy production GB of RAM and 15 GB of VRAM. forecasting and managing step. The Data Flow layers, com- The training process was configured with a maximum prising the Data Integration Layer and Climate Variables of 200 epochs and with the patience parameter in the early Broadcasting Layer, handle the acquisition, integration, and stopping mechanism set to 10 iterations. processing of heterogeneous data sources, providing reliable climate forecasts through advanced DL models. The Energy 4.2. Results Management layers, consisting of the Energy Production Forecasting Layer and Energy Flow Optimization Layer, To evaluate the comparative performance of the GRU and leverage these forecasts to optimize energy production and TCN models, we computed the Mean Absolute Error (MAE), distribution within the community through physical models the Coefficient of Determination (R²) and analyzed their and multi-agent reinforcement learning techniques. respective Taylor diagrams. The first step includes the Data Integration and Climate The MAE, reported in Equation (1), measures the aver- Variables Broadcasting layers, which are involved in the age absolute difference between the predicted values and acquisition, integration, and processing of heterogeneous the actual target values by giving equal weight to all errors, data sources to produce forecasts of climatic variables. The regardless of their direction (overestimates or underesti- data used in this step come from two different sources, as mates). previously introduced: real data, which belongs to physical 𝑛 weather stations, and MM5 model predictions. To assess 1 ∑︁ 𝑀 𝐴𝐸 = |𝑦𝑖 − 𝑦ˆ𝑖 | (1) this stage, the advanced deep learning models introduced 𝑛 𝑖=1 in Fig. 3 are adopted. These weather forecasts are then integrated into an effi- The Coefficient of Determination (CoD), calculated as de- cient prediction system for prosumers, using physical mod- fined in Equation (2), measures how well the model is able els and multi-agent RL to optimize energy production and to predict the variance of the data. The closer its value is to distribution. Our analysis compares two models for climate unity, the better the model fits the data. 𝑦¯ denotes the mean variable prediction: GRUs and TCNs. We evaluate their of the observed data, while 𝑦ˆ the predicted value. similarities, differences, and relative performance. These ar- ∑︀𝑛 chitectures were selected for their ability to track temporal (𝑦𝑖 − 𝑦ˆ𝑖 )2 𝑅2 = 1 − ∑︀𝑖=1 𝑛 (2) data evolution while identifying relevant patterns. 𝑖=1 (𝑦𝑖 − 𝑦¯ )2 These supervised learning models enhance the prediction As presented in [23, 24], the AI model manages to improve of key climatic variables - including GHI, temperature, pres- the forecasts returned by the MM5 RCM, which exhibits sure, and relative humidity. By combining RCMs forecasts both positive and negative deviations from the real data. In with actual observations as inputs, the models produce more this experiment, good performances are achieved for both accurate predictions and reduce the RCMs’ tendency to over- the GRUs and the TCNs, as shown in Tables 1 and 2 . or underestimate values during certain periods [23, 24]. Considering the Casaccia study (see Table 1), it is readibly observable that the GRU and TCN models applied to the MM5 model outperform it. The most impressive results Variable System 𝜏 = 144 (1 day) 𝜏 = 288 (2 days) 𝜏 = 432 (3 days) MAE R2 MAE R2 MAE R2 MM5 46.557 0.867 46.472 0.863 45.365 0.867 GHI GRU 39.140 0.902 39.863 0.892 37.907 0.944 TCN 37.755 0.906 38.778 0.902 37.626 0.903 MM5 2.389 0.841 2.388 0.840 2.361 0.844 Temperature GRU 1.232 0.948 1.366 0.940 1.313 0.942 TCN 1.252 0.947 1.329 0.940 1.363 0.937 MM5 173.116 0.863 173.264 0.857 170.598 0.867 Pressure GRU 90.423 0.954 90.062 0.948 90.151 0.952 TCN 77.956 0.966 85.129 0.956 81.097 0.961 MM5 10.720 0.571 10.735 0.570 10.715 0.581 Humidity GRU 8.313 0.733 8.498 0.724 8.480 0.728 TCN 8.371 0.733 8.775 0.706 8.654 0.716 Table 1 Comparison of performance metrics for Casaccia. Variable System 𝜏 = 144 (1 day) 𝜏 = 288 (2 days) 𝜏 = 432 (3 days) MAE R2 MAE R2 MAE R2 MM5 60.230 0.846 59.837 0.844 57.498 0.856 GHI GRU 52.595 0.887 53.566 0.882 53.105 0.878 TCN 50.282 0.898 53.846 0.933 53.471 0.884 MM5 60.757 0.977 64.548 0.975 64.392 0.977 Pressure GRU 70.594 0.964 72.816 0.968 95.708 0.954 TCN 67.796 0.967 70.835 0.971 74.626 0.969 MM5 10.353 0.680 10.158 0.685 9.982 0.697 Humidity GRU 7.794 0.817 8.239 0.794 9.049 0.750 TCN 8.088 0.797 8.196 0.786 7.933 0.799 Table 2 Comparison of performance metrics for Ottana. can be seen in atmospheric pressure forecasting, where the outperforms GRU, though both models maintain high accu- TCN reduce the error of over the 54.97%, 50.86% and 52.46% racy. The performance difference between the two locations with 1-day, 2-day and 3-day forecast horizon. At the same may stem from their different training data durations (three time, the CoD is raised by the 11.94%, 11.55% and 10.84%, years for Casaccia versus one year for Ottana). Similarly, respectively. Similarly, for the GHI parameter the TCNs both the superior performance of MM5 and TCN’s enhanced exhibit superior performance, although often comparable learning capabilities in the Ottana dataset likely reflect the to that of GRUs. The best results are achieved in the 1-day limited one-year training period. Future studies should in- forecast with a MAE decrease of 18.90% associated to the vestigate this relationship by testing model performance TCN, and in the 3-day forecast with an R² increase of 8.9% for across different time spans. the GRU. Regarding temperature and relative humidity, the Beyond standard R2 and MAE statistical measures, the GRU shows slightly better results than the TCN, although study incorporates Taylor diagrams as visual analytical tools their results remain very similar. to evaluate how well the method performs in handling mul- Concerning Ottana’s analysis reported in Table 2, the tiple variables simultaneously. These diagrams integrate MM5 model achieves better results than the ANNs regard- three key statistical measures into a single polar plot: the ing pressure forecasting; nevertheless, they still maintain Standard Deviation (𝜎), Correlation Coefficient (R), and Cen- remarkable effectiveness. In summary, a high degree of tered Root Mean Square Difference (RMSD). The reference correspondence is observed between the neural networks’ observations are positioned at the plot’s origin, while pre- forecasts and the experimental data, validating their predic- dicted values appear as points distributed across the diagram tive capabilities. The best MAE and R² performance can be based on their statistical properties. read in the humidity 1-day GRU prediction, with a reduc- The standard deviation functions as a measure of the vari- tion of 24.73% and an increase of 21% compared to the MM5 ability of the data, calculating how the values deviate from forecast. the mean. The correlation coefficient indicates the strength After having discussed the comparison between the neu- of the relationship between variables on a scale of −1 to ral networks and the MM5 model, the following section +1, where zero shows no relationship, positive values indi- presents a comparative assessment of GRUs and TCNs only. cate parallel movement and negative values suggest inverse It is important to emphasize that the two datasets include relationships. The root mean square deviation evaluates different time intervals: the Casaccia database covers three prediction accuracy by measuring the typical distance be- years, while the Ottana dataset only one. In both cases, the tween corresponding points in two datasets, with smaller TCN model outperforms the GRU in pressure forecast and values indicating better alignment. For any comparison be- it shows significant performance gains. In detail, for the tween simulated values (𝑓 ) and reference measurements Casaccia study, GRU and TCN models show comparable (𝑟), these metrics are calculated using established statistical performance across all parameters except atmospheric pres- formulations as detailed in [25]. sure. However, in the Ottana analysis, TCN consistently Figures 4 through 6 present the Taylor diagrams for Casac- (a) (b) Figure 4: Taylor diagrams for 1-day predictions referred to (a) Casaccia and (b) Ottana. cia and Ottana, corresponding to 1-day, 2-day, and 3-day approaches in enhancing weather forecasting accuracy, par- predictions, respectively. It is important to note that tem- ticularly through applying GRU and TCN architectures to perature analysis is limited to the Casaccia database, while improve MM5 model predictions. The comparative analysis the remaining climatic parameters are analyzed for both reveals several key findings that advance the field of weather locations. prediction and its applications in energy management. The For Casaccia, the Taylor diagram for 1-day predictions results show that both GRU and TCN models generally out- (see Fig. 4a) illustrates that GRU and TCN outperform MM5 perform the traditional MM5 model across multiple weather across all weather variables. Specifically, both GRU and parameters and time horizons. Notably, TCN demonstrated TCN achieve high correlation coefficients (above 0.9) for superior performance in atmospheric pressure forecasting, global solar radiation and temperature, positioning them achieving error reductions of up to 54.97% in one-day fore- close to the reference point and indicating strong predictive casts for the Casaccia dataset. While both neural network accuracy. In the case of atmospheric pressure, GRU and TCN architectures showed comparable effectiveness in predicting also perform well, with GRU slightly closer to the reference most parameters, TCN exhibited slightly stronger learning point. For relative humidity, TCN emerges as the most capabilities in scenarios with limited training data. accurate model, exhibiting the highest correlation and the An important finding emerged regarding the impact of closest match to observed data, while MM5 demonstrates training data duration on model performance. The contrast- the lowest correlation and the largest deviations across all ing results between Casaccia and Ottana suggest that the variables. Overall, TCN and GRU are identified as the most length of the training period significantly influences predic- reliable models for weather prediction in Casaccia. tion accuracy. This was particularly evident in the Ottana In Ottana (see Fig. 4b), the Taylor diagram highlights dataset, where MM5 maintained superior performance in the capabilities of GRU and TCN in predicting global solar pressure forecasting, highlighting the importance of com- radiation. For relative humidity, the models show mixed prehensive training data for neural network models. The results: one side achieves higher correlation coefficients and Taylor diagram analysis further validated these findings, lower RMSD, while the other side exhibits poorer standard demonstrating high correlation coefficients for both GRU deviation. In this case, MM5 performs better in terms of and TCN in predicting global solar radiation and temper- standard deviation. Finally, the RCM model surpasses the ature. This superior performance was maintained across AI models in estimating atmospheric pressure values. different prediction horizons (1-, 2-, and 3-day forecasts), The observed trends for each locality are consistently confirming the models’ reliability and stability. replicated in the 2-day (Fig. 5) and 3-day (Fig. 6) predic- These analyses serve as the foundation for the PRECEDE tions, underscoring the ability of both TCN and GRU to project’s next phase, which aims to extend these forecasting outperform traditional regional climate models in weather methodologies to cities in Emilia Romagna. The insights forecasting. gained from the Casaccia and Ottana studies will inform the implementation of these models across the region, sup- porting the project’s goal of enhancing renewable energy 5. Conclusions and Future management and community-based energy systems. More- Perspectives over, expanding the analysis to diverse climatic conditions and extended temporal scales will provide deeper insights Conducted within the framework of the PRECEDE project, into the robustness and adaptability of the proposed models. this study demonstrated the effectiveness of deep learning We would also like to test the Transformer model. (a) (b) Figure 5: Taylor diagrams for 2-day predictions referred to (a) Casaccia and (b) Ottana. (a) (b) Figure 6: Taylor diagrams for 3-day predictions referred to (a) Casaccia and (b) Ottana. Acknowledgments tax.html): During the preparation of this work, the author(s) used This work was supported by the PRECEDE project (CUP: X-GPT-4 and Gramby in order to: Grammar and spelling E93C22001100001), from the resources of the National Re- check. Further, the author(s) used X-AI-IMG for figures covery and Resilience Plan (NRRP), funded by the European 3 and 4 in order to: Generate images. After using these Union - NextGenerationEU. tool(s)/service(s), the author(s) reviewed and edited the con- tent as needed and take(s) full responsibility for the publica- tion’s content. Declaration on Generative AI Either: References The author(s) have not employed any Generative AI tools. [1] F. Dattola, P. Iaquinta, M. Iusi, D. Federico, R. Greco, Or (by using the activity taxonomy in ceur-ws.org/genai- M. Talerico, V. Coscarella, L. Legato, I. Pellegrino, S. Bergamaschi, et al., Precede: Climate and energy M. Ghahremani, F. Palmieri, Y. Liu, Tem- forecasts to support energy communities with deep poral convolutional neural (tcn) network for learning models, in: 2024 IEEE International Confer- an effective weather forecasting using time- ence on Big Data (BigData), IEEE, 2024, pp. 4650–4658. series data from the local weather station, [2] N. Benti, M. Chaka, A. Semie, Forecasting renew- Soft Computing 24 (2020) 16453–16482. URL: able energy generation with machine learning and https://doi.org/10.1007/s00500-020-04954-0. deep learning: Current advances and future prospects, doi:10.1007/s00500-020-04954-0. Sustainability 15 (2023). URL: https://www.mdpi.com/ [14] M. Alam, F. Al-Ismail, M. Hossain, S. Rahman, Ensem- 2071-1050/15/9/7087. ble machine-learning models for accurate prediction [3] X. Ren, X. Li, K. Ren, J. Song, Z. Xu, K. Deng, of solar irradiation in bangladesh, Processes 11 (2023). X. Wang, Deep learning-based weather predic- URL: https://www.mdpi.com/2227-9717/11/3/908. tion: A survey, Big Data Research 23 (2021). [15] A. Mellit, S. Kalogirou, Artificial intelligence URL: https://www.sciencedirect.com/science/article/ techniques for photovoltaic applications: A review, pii/S2214579620300460. Progress In Energy And Combustion Science 34 [4] J. Fan, L. Wu, F. Zhang, H. Cai, W. Zeng, X. Wang, (2008) 574–632. URL: https://www.sciencedirect.com/ H. Zou, Empirical and machine learning models for science/article/pii/S0360128508000026. predicting daily global solar radiation from sunshine [16] A. Yadav, S. Chandel, Solar radiation prediction duration: A review and case study in china, Renewable using artificial neural network techniques: A re- And Sustainable Energy Reviews 100 (2019) 186–212. view, Renewable And Sustainable Energy Reviews URL: https://www.sciencedirect.com/science/article/ 33 (2014) 772–781. URL: https://www.sciencedirect. pii/S1364032118307226. com/science/article/pii/S1364032113005959. [5] Q. Yang, C. Lee, M. Tippett, A long short-term mem- [17] A. Paniagua-Tineo, S. Salcedo-Sanz, C. Casanova- ory model for global rapid intensification prediction, Mateo, E. Ortiz-García, M. Cony, E. Hernández-Martín, Weather And Forecasting 35 (2020) 1203–1220. Prediction of daily maximum temperature using a sup- [6] M. Mohandes, S. Rehman, T. Halawani, Estima- port vector regression algorithm, Renewable Energy tion of global solar radiation using artificial neural 36 (2011) 3054–3060. URL: https://www.sciencedirect. networks, Renewable Energy 14 (1998) 179–184. com/science/article/pii/S0960148111001443. URL: https://www.sciencedirect.com/science/article/ [18] D. Kreuzer, M. Munz, S. Schlüter, Short-term tempera- pii/S0960148198000652. ture forecasts using a convolutional neural network [7] M. Ghadiri, A. Marjani, S. Mohammadinia, S. Shi- — an application to different weather stations in ger- razian, An insight into the estimation of rel- many, Machine Learning With Applications 2 (2020) ative humidity of air using artificial intelligence 100007. URL: https://www.sciencedirect.com/science/ schemes, Environment, Development And Sustain- article/pii/S2666827020300074. ability 23 (2021) 10194–10222. URL: https://doi.org/10. [19] G. Palma, L. Guiducci, M. Stentati, A. Rizzo, S. Pao- 1007/s10668-020-01053-w. letti, Reinforcement learning for energy community [8] S. Bai, J. Z. Kolter, V. Koltun, An empirical evaluation management: A european-scale study, Energies 17 of generic convolutional and recurrent networks for (2024). sequence modeling, arXiv preprint arXiv:1803.01271 [20] K. X. Perez, M. Baldea, T. F. Edgar, Integrated hvac (2018). management and optimal scheduling of smart appli- [9] H. Qin, Comparison of deep learning models on time ances for community peak load reduction, Energy and series forecasting: a case study of dissolved oxygen Buildings 123 (2016). prediction, arXiv preprint arXiv:1911.08414 (2019). [21] W. Jiatong, L. Li, Z. Jiangfeng, Deep reinforcement doi:10.48550/arXiv.1911.08414. learning for energy trading and load scheduling in [10] A. K. Shaikh, A. Nazir, N. Khalique, A. S. Shah, N. Ad- residential peer-to-peer energy trading market, Inter- hikari, A new approach to seasonal energy con- national Journal of Electrical Power & Energy Systems sumption forecasting using temporal convolutional 147 (2023). networks, Results in Engineering 19 (2023) 101296. [22] Y. Ye, Y. Tang, H. Wang, X.-P. Zhang, G. Strbac, A scal- URL: https://www.sciencedirect.com/science/article/ able privacy-preserving multi-agent deep reinforce- pii/S2590123023004231. doi:https://doi.org/10. ment learning approach for large-scale peer-to-peer 1016/j.rineng.2023.101296. transactive energy trading, IEEE Transactions on [11] J. Yu, X. Li, L. Yang, L. Li, Z. Huang, K. Shen, X. Yang, Smart Grid 12 (2021). Z. Xu, D. Zhang, S. Du, Deep learning models for pv [23] L. Caroprese, M. Pierantozzi, C. Lops, S. Montelpare, power forecasting: Review, Energies 17 (2024) 3973. Dl2f: A deep learning model for the local forecasting doi:10.3390/en17163973. of renewable sources, Computers & Industrial Engi- [12] R. Elmousaid, N. Drioui, R. Elgouri, H. Agueny, neering 187 (2024). URL: https://www.sciencedirect. Y. Adnani, Accurate short-term ghi forecast- com/science/article/pii/S0360835223008094. ing using a novel temporal convolutional network [24] C. Lops, M. Pierantozzi, L. Caroprese, S. Montelpare, model, e-Prime - Advances in Electrical Engi- A deep learning approach for climate parameter esti- neering, Electronics and Energy 9 (2024) 100667. mations and renewable energy sources, in: 2023 IEEE URL: https://www.sciencedirect.com/science/article/ International Conference on Big Data (BigData), 2023, pii/S277267112400247X. doi:https://doi.org/10. pp. 3942–3951. 1016/j.prime.2024.100667. [25] K. E. Taylor, Taylor diagram primer, Work Paper [13] P. Hewage, A. Behera, M. Trovati, E. Pereira, (2005) 1–4. doi:doi:10.1029/2000JD900719.