Weather forecasting, a critical function impacting numerous sectors—such as transportation, agriculture, energy, and public safety—relies on accurate, timely predictions of atmospheric conditions. Traditionally, meteorologists interpret raw data and produce forecasts in natural language, which can be a source-intensive process requiring significant expertise and time. Descriptive forecasts, unlike numerical or graphical data, offer comprehensible narratives on weather conditions, making them indispensable for a broad audience. However, crafting these narratives manually not only consumes resources but also introduces potential for human error, especially when forecasts cover multiple regions and varied weather patterns.
In response to this challenge, our study investigates tools and methodologies to create a predictive assistant that automates descriptive weather forecast generation. Using state-of-the-art NLP and ML techniques, we developed a system that recommends descriptive forecasts, integrating AI into the daily operations of meteorologists. Our assistant aims to streamline the forecasting process, allowing meteorologists to focus on high-level analytical tasks, while the assistant delivers accurate, contextually relevant descriptions based on raw meteorological data.
The need for automation in descriptive weather forecasting has driven advances in NLP and ML, enabling systems that process meteorological data and generate text-based predictions. Early approaches relied on rule-based systems, where domain experts predefined phrases and structures for automated forecasts. Although effective, these systems are labor-intensive to maintain, requiring continuous updates and adjustments by meteorologists, and are limited inflexibility across languages and regions.
Recent developments in NLP introduce more sophisticated models, such as embedding-based and transformer-based models, that analyze large datasets to select or create text that best describes specific atmospheric conditions. Notable applications include the FoG system in Canada, Arria NLG in the UK, and GaliWeather in Spain's Galicia region. However, these systems are highly specialized, constrained by language and geographic context, and, in many cases, rely on predefined phrases that limit adaptability. While machine translation can extend some flexibility, localized geographic and climatic references pose challenges for broad, scalable application.
To address these limitations, our study integrates NLP with models capable of handling Portuguese text and adaptable to Portuguese geographical and meteorological nuances. We leverage recent advancements in transformers and topic modeling, aiming to create an assistant that adapts to Portuguese forecasting needs and supports automated forecast production across multiple contexts.
The technology stack for this study involves NLP frameworks, ML models, and specific libraries designed to manage large datasets and produce high-quality text output in natural language. Key technologies include:
BERT (Bidirectional Encoder Representations from Transformers): A transformer-based model that supports language understanding tasks, which we adapted for Portuguese using multilingual variations. BERT enables us to generate detailed, semantically rich embeddings of forecast descriptions.
BERTopic: This model allows for topic modeling by embedding text and clustering similar phrases. In our study, BERTopic aids in identifying and organizing key atmospheric events within forecast data, creating descriptive labels for various weather phenomena.
Zero-Shot Classification (ZSC): This technique facilitates classifying unseen data, allowing our model to identify weather events without prior training for each specific condition. ZSC helps verify that predicted labels align with relevant atmospheric events, even when data is limited.
Sentence Transformers and Hugging Face Models: Used to produce high-dimensional embeddings of forecast data, these models help us measure semantic similarity between text snippets, crucial for classifying and grouping forecast descriptions.
UMAP (Uniform Manifold Approximation and Projection) and HDBSCAN (Hierarchical Density-Based Spatial Clustering of Applications with Noise): Used in dimensionality reduction and clustering, these algorithms support the BERTopic model by enabling efficient grouping of semantically similar texts.
NLTK (Natural Language Toolkit): Employed for text pre-processing tasks, such as tokenization and removal of common stop words in Portuguese, enhancing model focus on meteorologically relevant terms.
c-TF-IDF (Class Term Frequency–Inverse Document Frequency): This modified TF-IDF approach helps identify keywords within clustered texts, providing essential descriptors for different atmospheric conditions.
Study Details
The primary goal of our study is to create an NLP-driven assistant that accurately generates descriptive weather forecasts in Portuguese, allowing meteorologists to shift their focus to complex analytical tasks by reducing the time and labor needed for routine text-based forecast generation. This assistant is a core component of a larger initiative to develop a “Virtual Meteorologist” system, designed to integrate numerical data analysis with descriptive forecast production.
The methodology centers on two main objectives:
- Classifying and Labeling Weather Descriptions: Using NLP models to categorize weather forecasts based on distinct meteorological events, such as heavy rain, temperature drops, or strong winds.
- Generating Contextually Accurate Recommendations: Developing a recommendation system to select the best forecast text based on real-time or anticipated weather conditions.
To achieve these goals, we used a dataset provided by the Portuguese Institute for Sea and Atmosphere (IPMA), comprising of manually crafted descriptive forecasts spanning various regions of Portugal. This dataset served as the foundation for training and validating our NLP models, with particular attention to seasonal variations and geographic nuances in weather patterns.
Model Training and Data Processing
Our study’s approach involved several stages:
Data Pre-Processing: We started by pre-processing the forecast texts for major regions such as the North, Center, South, Lisbon Metropolitan Area, and Porto Metropolitan Area. Texts were segmented into sentences to facilitate finer-grained analysis of specific meteorological events. Using NLTK, we removed common Portuguese stop words, focusing on meteorologically significant terms to improve classification accuracy.
Topic Modeling with BERTopic: We employed the BERTopic model, which uses embedding and clustering to identify recurring themes within text data. Embedding models from Hugging Face and SentenceTransformers provided rich semantic representations, enabling us to group forecast descriptions into meaningful topics. By applying dimensionality reduction with UMAP and clustering with HDBSCAN, BERTopic effectively grouped similar forecast descriptions. The model was optimized by adjusting hyperparameters, such as the number of neighbors and dimensions, to balance accuracy and computational efficiency.
Zero-Shot Classification for Label Mapping: To align forecast topics with meteorological event labels, we applied Zero-Shot Classification (ZSC), allowing the system to identify classes within text data without explicit prior training on each event type. This step was crucial for validating topic clusters and ensuring they accurately represented the target labels, based on IPMA’s established categorization of atmospheric conditions. The ZSC approach also allowed flexibility in adjusting or expanding the forecast system to accommodate new weather events or rare occurrences.
Forecast Recommendation System: The final stage involved developing a recommendation algorithm that matches descriptive forecasts to new or predicted weather conditions. Using c-TF-IDF, we generated topic/label-based descriptors for each forecast text. The algorithm assesses the cosine similarity between these descriptors and current weather conditions, selecting the most contextually appropriate forecast text.
Findings and Results
The results show that our automated assistant performs comparably to manual methods, producing forecasts with an average accuracy of 45% when benchmarked against a manual “gold standard” classification. The assistant excelled in identifying specific events, such as “partly cloudy skies,” “ice,” “snow,” and “strong winds,” achieving over 70% accuracy for these labels. This high level of accuracy demonstrates the system’s robustness in capturing essential details for more predictable meteorological events, even with the complexities of natural language processing.
Our testing revealed several insights:
- Seasonal Pattern Recognition: The recommendation system reliably captured seasonal patterns, aligning forecast suggestions closely with corresponding seasonal weather. Texts for winter months, for instance, frequently included descriptors for snow or ice, mirroring expected seasonal shifts in weather phenomena.
- Consistency in Forecast Descriptors: By comparing similarity distributions across recommended and actual forecast texts, we observed a stronger dissimilarity in label-based text descriptors(4.038 natural units of Kullback-Leibler divergence) vs text dissimilarity (0.181 natural units),well segregated text descriptions even when the underlying texts vary slightly. This outcome validates the system’s ability to deliver relevant, contextually rich recommendations.
- Challenges in Text Similarity: While label-based dissimilarity was strong, purely text-based similarity posed challenges due to limited lexical variety within forecast descriptions. Variations in phrasing for similar conditions (e.g., “sharp temperature drop” vs. “significant cooling”) complicated direct text-based matching, highlighting the need for further refinement in text embeddings for finer nuances in forecast language.
Technically, our study demonstrates the feasibility of automating descriptive weather forecasting using NLP and ML. The assistant’s classification and recommendation capabilities align closely with manual efforts, achieving a practical level of accuracy and reducing the time meteorologists spend on routine descriptive tasks. Integrating such automation into existing systems enhances operational efficiency and minimizes the potential for human error, offering a scalable solution for IPMA and other meteorological institutions.
