Accurate Long-term Air Temperature Prediction with a Fusion of Artificial Intelligence and Data Reduction Techniques

Dusan Fister, Jorge Pérez-Aracil, César Peláez-Rodríguez, Javier Del Ser, Sancho Salcedo-SanzarXiv 2022 · OpenReview

TL;DR

This paper introduces three novel AI frameworks that fuse deep learning and machine learning with data reduction techniques to achieve accurate long-term air temperature predictions for summer months. By carefully selecting relevant data and employing advanced models, the research demonstrates a significant improvement in forecasting extreme weather events like heatwaves, which is crucial for climate change adaptation and renewable energy planning.

What Problem Does This Paper Address?

Accurately predicting long-term air temperature, especially during summer months, is a critical challenge with significant implications for climate change adaptation, risk management, and energy planning. This is particularly true for predicting extreme events like heatwaves, which can have devastating consequences. Traditional forecasting methods often struggle with the complexity and long time horizons involved. This paper aims to bridge this gap by developing and evaluating advanced AI-based frameworks that can effectively capture the intricate patterns in climate data to provide more reliable long-term temperature forecasts, even for extreme events.

Methods

The authors propose three distinct AI frameworks, all leveraging data reduction and feature selection techniques:

  1. Convolutional Neural Networks (CNNs): This framework utilizes 3D CNNs for "video-to-image" translation, treating sequences of climate data as video frames. This approach effectively processes spatio-temporal data to capture complex patterns.
  2. Machine Learning (ML) Methods: This framework explores several established ML algorithms, including linear regression, Lasso, polynomial regression, AdaBoost, decision trees, and Random Forests. These methods operate on time-series data rather than images.
  3. Recurrence Plot (RP) with CNN: This innovative approach combines Recurrence Plots (RPs) with CNNs. RPs transform time-series data into images, preserving temporal structure, which are then fed into a CNN. Both classic (analogue) and binarized RPs are investigated.

A crucial preprocessing step across all frameworks is Geographic Area Selection (GAS), where correlation analysis is used to identify regions with the highest correlation to the target temperature, thereby focusing the analysis on the most relevant geographical areas. Additionally, Exhaustive Feature Search (EFS) is employed to identify the optimal subset of predictor variables for each model, further enhancing prediction accuracy and efficiency.

Datasets

The study utilizes the ERA5 Reanalysis dataset, a comprehensive global atmospheric reanalysis dataset maintained by the European Centre for Medium-Range Weather Forecasts (ECMWF). This dataset provides hourly meteorological data from January 1950 to December 2021, covering a wide geographical area. The data is downsampled to a semi-monthly (fortnightly) basis for analysis. The study focuses on two specific locations: Paris, France, and Córdoba, Spain.

Key Insights and Ideas

  • The power of data reduction: Identifying and focusing on geographically relevant areas (GAS) and selecting the most informative features (EFS) significantly improves prediction accuracy, demonstrating that not all data is equally valuable for forecasting.
  • Temporal structure matters: The Recurrence Plot (RP) technique effectively captures the temporal dynamics of climate data, transforming it into an image format that CNNs can process, leading to improved performance, especially for the RP+CNN and RP+CNN+BIN models.
  • No one-size-fits-all: The "best" model varies depending on the prediction horizon (shorter vs. longer) and the geographical location, highlighting the need for adaptable and context-aware forecasting approaches.
  • Extreme events are predictable: The proposed AI frameworks show promise in capturing signals of extreme events like heatwaves, which is a significant advancement over traditional methods.

Key Results

The study found that the proposed AI frameworks, particularly the RP+CNN and RP+CNN+BIN approaches, demonstrated strong prediction skills for long-term air temperature in both Paris and Córdoba.

  • Paris: For shorter prediction horizons (T1), the Poly model showed good performance, while for longer horizons (T2), RP+CNN+BIN was the most promising.
  • Córdoba: The RP+CNN+BIN model consistently outperformed others for shorter horizons (T1), and RP+CNN for longer horizons (T2). The study noted that Córdoba generally showed better forecastability than Paris.
  • General Trend: The more complex the modeling method, the better the forecast for prolonged time horizons.
  • EFS Impact: The Exhaustive Feature Search significantly reduced the number of predictor variables, with some models using as few as two variables while maintaining high accuracy.

Limitations and Future Work

  • Data Availability: While ERA5 data is extensive, the authors acknowledge that for many extreme events, the historical data might be limited, posing a challenge for training complex DL models.
  • Model Specificity: The study found that no single model is universally best for all geographical areas and prediction horizons, suggesting the need for adaptive or ensemble approaches.
  • Future Work: The authors suggest several avenues for future research:
  • Expanding the CNN model to output multiple channels (e.g., wind, soil water) instead of just air temperature.
  • Exploring data augmentation techniques to improve model stability with increased complexity.
  • Investigating auto-encoders for image compression and reconstruction.
  • Utilizing a larger climate data span (e.g., including January-April and September-December) to better capture climate trends.
  • Developing a universal model that can accurately forecast for all geographical areas.

Summary updated

Abstract

In this paper three customised Artificial Intelligence (AI) frameworks, considering Deep Learning (convolutional neural networks), Machine Learning algorithms and data reduction techniques are proposed, for a problem of long-term summer air temperature prediction. Specifically, the prediction of average air temperature in the first and second August fortnights, using input data from previous months, at two different locations, Paris (France) and C\'ordoba (Spain), is considered. The target variable, mainly in the first August fortnight, can contain signals of extreme events such as heatwaves, like the mega-heatwave of 2003, which affected France and the Iberian Peninsula. Thus, an accurate prediction of long-term air temperature may be valuable also for different problems related to climate change, such as attribution of extreme events, and in other problems related to renewable energy. The analysis carried out this work is based on Reanalysis data, which are first processed by a correlation analysis among different prediction variables and the target (average air temperature in August first and second fortnights). An area with the largest correlation is located, and the variables within, after a feature selection process, are the input of different deep learning and ML algorithms. The experiments carried out show a very good prediction skill in the three proposed AI frameworks, both in Paris and C\'ordoba regions.
BibTeX
@misc{fister2022accurate,
  author = {Dusan Fister and Jorge Pérez-Aracil and César Peláez-Rodríguez and Javier Del Ser and Sancho Salcedo-Sanz},
  title = {{Accurate Long-term Air Temperature Prediction with a Fusion of Artificial Intelligence and Data Reduction Techniques}},
  year = {2022},
  eprint = {2209.15424},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2209.15424},
}
Figure