AI Weather Forecasting: Educational Insights from a FE Competition on Meteorological LLMs

Photo by Jin Zhang / Pexels

---

On the day of the Shanghai FE competition, I was in the lab guiding students to debug a small project using a Raspberry Pi and a camera—they wanted to predict if it would rain in ten minutes by analyzing sky photos. Naturally, they failed repeatedly, with the model oscillating wildly between cloudy and sunny boundaries. I closed the terminal and showed them the news about Envision Tianqi Meteorological Large Model assisting the FE team. The kids' eyes lit up: "Teacher, how did they do that? Can we do it too?" This made me realize that this news piece is actually an excellent teaching slice for AI education.

Conclusion first: The success of the Envision Tianqi Meteorological Large Model at the FE Shanghai round wasn't just a simple tech flex, but a typical example of AI providing "decision support" in the physical world. It used minute-level ultra-short-term rainfall prediction to convert abstract meteorological data into actionable tire strategies and energy management plans for the team. For middle school students, this case connects the complete loop of "Data Collection—Model Training—Real-time Inference—Decision Feedback," with a clear tech stack, making it very suitable as a reference framework for science innovation projects.

---

1. Technical Breakdown: From "Large Models" to "Small Scenarios"

The news mentions the "Envision Tianqi" meteorological large model, but what truly played a role was its ultra-short-term precipitation prediction capability. This step didn't involve retraining a massive model with tens of billions of parameters, but rather fine-tuning an existing large model specifically for the task of rainfall prediction.

[!example] Simplified Analogy of the Tech Route

It's like giving a student a model that already knows how to identify cats and dogs, then showing it only pictures of "cats" and asking it to predict where the cat will jump next. The model doesn't need to relearn the world; it just needs to "focus" and "predict" within a specific scenario.

Specifically, this type of prediction relies on time-series radar echo extrapolation technology. Traditional methods use optical flow or ConvLSTM, while Envision Tianqi likely combines Transformer architecture's time-series modeling capabilities, taking sequences of radar echo images from the past few dozen minutes as input and outputting rainfall intensity maps for the next 30 minutes. Its accuracy reaches the minute level, which is crucial for the FE team at the "unpredictable weather" Shanghai round—tire selection (dry vs. wet tires) and energy recovery strategies (changes in braking points on slippery surfaces) require decisions within minutes.

# Simplified structure of an ultra-short-term rainfall prediction model (pseudocode for teaching)
class ShortTermRainPrediction(nn.Module):
    def __init__(self, backbone='4d_weather_former'):
        # Use a pre-trained meteorological large model backbone
        self.encoder = load_pretrained_weather_encoder()
        # Add a lightweight prediction head specifically for future 30-min rainfall prediction
        self.pred_head = nn.Sequential(
            nn.Conv3d(256, 128, kernel_size=3, padding=1),
            nn.GRU(input_size=128, hidden_size=64, num_layers=2),
            nn.Conv2d(64, 1, kernel_size=1)  # Output rainfall intensity map
        )
    
    def forward(self, radar_sequence):
        # radar_sequence: shape (batch, time, channels, H, W)
        features = self.encoder(radar_sequence)
        rain_pred = self.pred_head(features)
        return rain_pred  # Minute-by-minute rainfall intensity for the next 30 mins

Key Point: This isn't an "all-purpose" weather model, but a domain-specific small model. It consumes far fewer computational resources than training a general large model, yet its accuracy and speed surpass general models. This "Large Model Foundation + Specific Fine-tuning" approach is a key paradigm for current AI implementation and is something middle school students can understand and try to replicate—they can train a simplified prediction model using small-scale radar datasets (like public CASA radar data).

---

2. Educational Value: From "Watching the Show" to "Understanding the Craft"

This case has three levels of educational value in AI education:

1. Intuitive Presentation of Data-Driven Decision Making: The team doesn't directly use the "rainfall probability" number but converts the prediction results into two concrete actions: "Tire Strategy" and "Energy Management." This shows students that AI outputs must be combined with business logic to have value. A common mistake in student science projects is pursuing model accuracy while ignoring how the model is used.

2. Conceptual Bridge for Time-Series Prediction: Weather forecasting is one of the most understandable time-series prediction tasks for students. Compared to image classification (cat/dog recognition) familiar to students, the question "Will it rain next?" naturally involves the time dimension, allowing for the introduction of concepts like RNN, LSTM, and Transformer. And instead of jumping straight into mathematical formulas, use analogies like "The model looks at past frames, predicting the next frame like watching a cartoon."

3. Importance of Small Parameter Counts: The AFE team needs to deploy edge computing or real-time communication on the race car, so the model must be lightweight. This brings up concepts like "Model Compression" and "Knowledge Distillation." Students can experience firsthand: A large ResNet-50 model won't run on a Raspberry Pi, but after pruning and quantization, a smaller model with the same accuracy runs smoothly.

[!info] Learning Curve Assessment

This case is suitable as an AI science innovation project theme for upper middle school or high school students. Prerequisites include:

- Basic Python programming (ability to handle arrays and images)

- Understanding of basic machine learning concepts (training, testing, overfitting)

- Willingness to touch Numpy and PIL for image sequence operations

Advanced direction: Understanding radar data formats and attention mechanisms in time-series models.

Original link: https://www.tmtpost.com/8062394.html