Application of neural networks for meteorological time series analysis and forecasting: from RNN to LSTM

Area of interest: Investigation of capabilities of recurrent neural networks, in particular RNN and LSTM models, for solving the problem of meteorological time series forecasting. The potential benefits of the introduction of such technologies are analysed in the context of global challenges: climate change, food safety and effective resource management.

Relevance of the topic
The modern world is experiencing an unprecedented growth in data volume, including meteorological data, which is extremely important for decision-making in agriculture, energy, city planning, and disaster prevention. Technological advances enable the use of complex machine learning algorithms for analysing and forecasting weather conditions. One of the key trends has been the integration of neural networks capable of processing time series and creating detailed scenario forecasts.
Improvements in the accuracy of meteorological forecasts directly affect socioeconomic stability, particularly in countries of the Global South and East, where agriculture and water resources require careful management. In this context, neural networks pave the way to increased planning efficiency and reduction of climate change risks. Based on this, the problem of time series forecasting using RNN and LSTM is important not simply fr om the technical standpoint, but also fr om a global perspective, as it favours a transition to more sustainable forms of economic activity and contributes to the development of a new platform for global growth.

Research hypothesis and goals

Hypothesis. More advanced neural network architectures (LSTM), relative to classical RNN, provide greater accuracy of meteorological time series forecasting, especially under conditions of complex seasonality and temperature extremes.

Goal. To compare several recurrent network variants (RNN, Keras-based and PyTorch-based LSTM) in terms of prediction accuracy and stability.

Analysis and statistical data. The experiment used weather data at threehour time intervals (city of Tashkent, OpenWeatherMap). After collection and preprocessing (normalization to the range [0;1]), the models were trained on 80% of the dataset, and the remaining 20% were used for testing. The key accuracy parameter was RMSE (root-mean-square error).

1. RNN: RMSE ≈ 4.13. The least satisfactory results were obtained when forecasting complex seasonal variations, especially during extreme heat events (Fig. 1).
2. LSTM (Keras): RMSE ≈ 3.53. The model confidently forecasts weather conditions, but accuracy is degraded at extreme temperatures.
3. LSTM (PyTorch): RMSE ≈ 3.56. Comparable to Keras results, which speaks
to the stability of the LSTM-based approach for time series (Fig. 2 and Fig. 3).

Forecasts. The results warrant the following expectations:
• An increasingly prominent role of LSTM models (and their modifications like GRU) in meteorology and related fields;
• Reduction of economic losses due to weather forecasting errors (losses in agriculture, in the power industry, etc.);
• Growth of investment attractiveness and emergence of new technological solutions in countries of the Global South and East, wh ere it is especially important to take climate change into account and implement prudent resource management.

Methodology of data collection and analysis
• Data source: OpenWeatherMap with an open API was used for building the dataset. Weather data for the city of Tashkent were loaded in JSON format with a 3-hour time step.
• Structuring: Field extraction (temperature, humidity, precipitation, wind speed, etc) to table weather_history_test.

Wh ere country, city and day were taken from date and request nodes, and the other fields were taken from the hourly list of measured weather parameters. Each of eight measurements throughout the 24-hour period was associated with a name of a city.
• Preprocessing: data cleanup and normalization using the pandas library (and if necessary, SQLite3), which converted the data to the [0;1] range. Description of the investigated architectures
1. RNN (SimpleRNN)
→ Three layers: input, hidden (SimpleRNN) and output (Dense).
→ Optimizer: SGD with a 0.7 learning rate.
→ Loss function: mean squared error (MSE).
→ Result check every 10 epochs.
2. LSTM (Keras)
→ Includes the input layer, the LSTM layer with 50 cells and the output layer.
→ Optimizer: Adam, loss function: MSE.
→ Training on 80% of the dataset, the remaining 20% are allocated for testing.
3. LSTM (PyTorch)
→ A similar LSTM architecture but based on PyTorch.
→ DataLoader objects are used for serial loading of data batches.
→ RMSE (root-mean-square error) is the key accuracy indicator.

Test results
• RNN: RMSE ≈ 4.13, the least satisfactory results when forecasting complex seasonal variations, especially extreme summer temperatures.
• LSTM (Keras): RMSE ≈ 3.53. The model confidently forecasts weather conditions, but accuracy is degraded during extreme heat.
• LSTM (PyTorch): RMSE ≈ 3.56. Comparable to the results of the Keras version, which confirms the stability of the LSTM-based approach to time series. The overall trend shows that LSTM-based networks are better than RNN at handling “remote dependencies” in time series, especially during periods with extreme temperatures.

Economic effect
1. Loss prevention: more accurate weather forecasts help reduce losses in agriculture, logistics and the power industry, which, in turn, improves the financial stability of regions.
2. Growth of investment attractiveness: The implementation of hi-tech solutions improves the competitive position of countries of the Global South and East and encourages the arrival of investors and development of AI startups.

Social effect
1. Enhanced safety: Accurate forecasts help warn the population of natural disasters in a timely manner. This is especially important in coastal and mountainous regions that are at risk of floods or hurricanes.
2. Closing the digital divide: The training of professionals and dissemination of AI technologies stimulate the development of educational programmes and build new competencies among young people and leaders.
3. Building the tech ecosystem: The development of neural networks in meteorology pulls related sectors along, from IoT devices to platform solutions for urban planning and big data analysis.
Therefore, the large-scale application of neural networks for weather data analysis serve as a catalyst for global growth, combining economic benefits with social stability.

The results confirm that:
1. RNN is capable of solving basic forecasting tasks, but it is visibly inferior to LSTM models in terms of accuracy.
2. LSTM (Keras) and LSTM (PyTorch) demonstrate high efficiency, although they remain vulnerable to temperature extremes.
3. The optimum choice of architecture for meteorological time series most commonly tends towards LSTM, taking better of long series and more stable error parameters (RMSE) into account. In the long term, the application of neural networks for weather forecasting will grow, encompassing more and more regions and scenarios (agriculture, water resource management, power industry). This will ensure accelerated technology development in countries with a vulnerable infrastructure, create incentives for scientific and commercial collaboration, and contribute to global stability in the face of climate challenges.

References
1. Smirnov, S. M., Kuznetsov, K. K., & Novikov, N. N. (2023). Global Warming and its
Consequences for Humanity. Mirovaya politika, 3, 34–48. (in Russian)
2. https://openweathermap.org/api – open API for receiving weather data at different levels of
detail.