Using LSTM Deep Learning to Predict Sudden Wind and Waves in Bays: A Real-Time Meteorological Early Warning System for Barnegat Bay, New Jersey

This project combines LSTM neural networks, linear regression, and naive prediction models to perform time-series forecasting of sudden onshore wind events and upwelling conditions in Barnegat Bay, New Jersey, aiming to provide one-hour-ahead wind condition warnings for water sports enthusiasts. The project integrates multi-source monitoring data, designs an upwelling detection algorithm based on domain knowledge, compares the performance of different models, and finally develops a real-time early warning platform.

Barnegat Bay is a popular water activity destination in summer, but sudden strong onshore winds are often accompanied by upwelling, forming dangerous waves. Traditional regional-scale forecasts struggle to capture small-scale changes in the bay, leading to a lack of timely warnings for water sports participants and potential safety risks. Core project question: Can a local LSTM model trained on hourly observation data provide more accurate location-specific one-hour-ahead warnings?

Forecast Objectives: For the Stafford mainland weather station, predict wind speed, gust speed, wind direction, onshore indicator, and upwelling indicator for the next hour. Data Sources: Integrate hourly data from 4 monitoring stations (mainland/Long Beach Island weather stations, bay/ocean buoys) from June to September 2025. Feature Engineering: Standardize numerical features, encode wind direction as integers from 0 to 15, unify temperature units to Celsius. Upwelling Detection: Determine based on wind direction filtering (SE/SSE/S/W) and offshore temperature difference (past 6 hours vs. 48 hours ≤ -3°C and lasting for 6 hours). Model Architecture: Compare LSTM (24-hour sliding window), linear regression, and naive persistence models.

Compare model performance using MAE metrics:

Target Variable	Linear Regression MAE	Naive Model MAE	LSTM Status
Wind Speed	0.56	1.13	Collapsed
Gust Speed	0.96	1.17	Collapsed
Wind Direction	3.92	1.17	Collapsed
Onshore Indicator	0.55	0.17	Collapsed
Upwelling Indicator	Collapsed	Collapsed	Collapsed
The onshore indicator derived from wind direction achieved an MAE of 0.13, which was the best performance. LSTM collapsed due to small data scale, class imbalance (rare upwelling events), and other factors.

The project developed multiple application forms: Online Platforms (main site: wind.krupamlab.com, backup site: barnegatbaywindprediction.streamlit.app) that pull data in real time for forecasting; Local Deployment (desktop/terminal scripts, executable programs, lightweight Raspberry Pi version); planning to develop seasonal model groups to achieve year-round forecasting.

Insights: Not all tasks require complex models; in this scenario, linear regression and naive models are more stable, due to factors such as small data scale (only 4 months of summer data), strong linear correlation between meteorological variables, effective persistence rules for short-term forecasting, and noise from multi-source data affecting complex models. Limitations: Upwelling prediction failed due to rare samples; the model was only trained in summer, and generalization to other seasons has not been verified.

The project demonstrates a complete meteorological forecasting workflow from data collection, feature engineering, domain knowledge integration to model training and deployment. Although LSTM did not meet expectations, the finding that "naive models are better than complex models" is of great value. Future directions: Collect longer-term data to re-evaluate LSTM; adopt class imbalance handling to solve upwelling prediction; expand the forecast window to provide earlier warnings.