Predicting Hurricane Trajectories using a Recurrent Neural Network

Motivation

• Hurricanes can cause severe economic disasters and loss of life, making accurate trajectory prediction important for evacuation and preparation.The passage cites Hurricane Betsy’s $1.5 billion property damage and Hurricane Katrina’s 853 recorded fatalities.Source: Introduction · p. 1
• Most current hurricane trajectory forecast methods are statistical, but atmospheric systems are complex and nonlinear.Source: Introduction · p. 1
• RNN weight modification enables learning of intricate dynamic temporal behaviors relevant to hurricane forecasting.Source: Introduction · p. 1

Contribution

• The paper proposes a fully connected recurrent neural network with a grid model for hurricane trajectory forecasts.The authors describe this combination as, to their knowledge, the first of its kind for hurricane trajectory forecasting.Source: Introduction · p. 1

Existing Forecast Models

• The NHC uses dynamical, statistical, statistical-dynamical, and ensemble or consensus models for hurricane path prediction.Source: Related Work · p. 1
• Dynamical models simulate atmospheric motion with high computational requirements and can generate detailed storm-track and meteorological outputs.The GFDL example uses nested movable meshes, vortex replacement, and a 72-hour prediction process.Source: Related Work · pp. 1–2
• Statistical models use historical storm features and spatially averaged data to simulate possible trajectory paths while applying out-of-sample validation to limit overfitting.Source: Related Work · p. 2
• Consensus models combine forecasts, physical parameters, or initial conditions and are more accurate than individual components on average.Source: Related Work · p. 2

Prior RNN Approaches

• Prior sparse RNN work uses historical dynamics, genetic-algorithm weight optimization, and dynamic time warping, but does not accommodate nonmonotonic hurricane paths.Source: Related Work · p. 2

Motivation for RNNs

• The paper motivates RNNs because atmospheric nonlinearity, increasing data volume, and limited use of observations constrain linear or sparse approaches.Source: Related Work · p. 2

RNN Foundations

• RNNs are nonlinear dynamical models used to represent complex dynamical or sequential relationships between variables.Source: Recurrent Neural Networks · p. 2

RNN Architecture

• A deep recurrent network arranges hidden-state vectors across time steps and network depth.The architecture uses t = 1, ..., T for total learning time and l = 1, ..., L for network depth.Source: Recurrent Neural Networks · p. 2
• Figure 1 depicts a simple recurrent neural network architecture with depth two.Source: Figure 1 · p. 2, Recurrent Neural Networks · p. 2

Output Mapping

• The output vector is produced by applying an activation function to the final hidden-state vector through an output weight matrix.Source: Recurrent Neural Networks · p. 2

Hidden-State Recurrence

• The hidden-state recurrence combines the previous hidden state and current input through learned weight matrices and an activation function.The recurrent weights model dynamic connections among hidden states and support latent nonlinear interactions.Source: Recurrent Neural Networks · pp. 2–3

Trajectory Context

• The paper includes hurricane trajectory examples and identifies largest-distance-traveled behavior as part of the modeled trajectory context.Source: Figure 2 · p. 3, Recurrent Neural Networks · p. 3

Dataset

• The dataset contains Atlantic hurricanes and tropical storms from 1920 to 2012 with six-hourly locations, wind speeds, and minimum central pressures.Locations are recorded in latitude and longitude tenths of degrees, and wind speeds are measured in knots.Source: Unisys Weather Atlantic Hurricane Data · p. 3

Feature Construction

• Distance traveled and travel angle are derived from successive latitude-longitude observations to provide relative rather than absolute parameters.The authors state that relative variables support learning for unseen paths.Source: Unisys Weather Atlantic Hurricane Data · p. 3
• The largest and smallest example hurricanes traveled approximately 6394.7 miles and 86.5 miles, respectively.Source: Unisys Weather Atlantic Hurricane Data · p. 3

Feature Analysis

• The number of recorded data points and total track distance have a Pearson correlation coefficient of 0.739.The authors use this relationship to motivate continuously supplied direction and angle information.Source: Unisys Weather Atlantic Hurricane Data · p. 3

Preprocessing

• Direction and travel-angle features reject normality under the Anderson-Darling test at significance level 0.05 and are normalized before modeling.Source: Unisys Weather Atlantic Hurricane Data · pp. 3–4
• Normalization uses Z = (X−µ)/σ, where X is the original value, Z the normalized value, and µ and σ the historical-path mean and standard deviation.Source: Unisys Weather Atlantic Hurricane Data · p. 4

Grid Model

• The fully connected RNN learns hurricane trajectory behavior across grid locations while accumulating historical information about nonlinear atmospheric dynamics.The grid model represents movement from one grid location to another and updates weight matrices to capture unobservable hurricane states.Source: Model and Implementation · p. 4, Grid Model · p. 4
• A refined grid over recorded latitude and longitude points reduces truncation errors while preserving small-scale trajectory features.The grid is intended to support a larger-scale model that more accurately encapsulates small-scale behavior.Source: Grid Model · p. 4

Implementation

• The model uses 7,256 one-degree grid blocks for 13,131 valid data points from 539 hurricane or tropical-storm trajectories.The number of grid blocks was selected in relation to the available training data.Source: Grid Boundaries · p. 4
• Dropout was set to 0.1 through cross-validation, randomly ignoring 10% of each input during training to reduce overfitting.Dropout is used as a regularization hyperparameter to prevent co-adaptation to the training set.Source: Dropout · p. 4

Network Architecture

• The network uses three hidden layers, each containing an LSTM cell, to model trajectory complexity without overfitting.The paper motivates three hidden layers as a balance between capturing complexity and avoiding overfitting.Source: Hidden State Vectors · pp. 4–5, Long Short-Term Memory Cell · p. 4
• The input data tuple contains wind speed, latitude and longitude coordinates, travel direction, distance, and grid identification number.The architecture contains five total layers: an input layer, three hidden layers, and an output layer.Source: Network Architecture and Implementation · p. 5

Implementation

• The output uses an LSTM unit, dropout, a dense layer, and a tanh activation that produces values between [-1, 1] for movement in all directions.The model was implemented in Keras using a sequential recurrent neural network structure.Source: Network Architecture and Implementation · p. 5

Trajectory predictions

• The grid-based RNN was evaluated on randomly selected, long-duration hurricanes from the upper half of the testing set.Figure 5 compares predicted grid locations with recorded locations for ALEX, DELTA, SANDY, HORTENSE, IVAN, and PALOMA.Source: Forecast Results · p. 5

Comparison with sparse RNN

• The model achieved lower mean absolute errors than the sparse RNN approach, whose training excluded hurricanes with loops.The sparse RNN therefore could predict only non-looping trajectories, whereas the grid-based comparison included dynamic hurricane paths.Source: Forecast Results · pp. 5–6

NHC comparison

• 48-hour forecasts were the primary comparison target because they are most relevant to disaster management and citizen preparedness.The model also produced forecasts ranging from 6 to 120 hours.Source: Forecast Results · p. 6
• The grid-based RNN outperformed currently employed NHC techniques in the 48-hour comparison, while both methods exceeded the GPRA requirement.The refined grid reduces truncation errors and captures small-scale features more accurately, allowing the RNN to learn relative movement between locations.Source: Forecast Results · p. 6, Figure 6 · p. 7

Computational performance

• The model completed training in 200 seconds, compared with the hours required by statistical-dynamical models for a single prediction.A trained neural network can make predictions instantly, which the paper identifies as important for up-to-date forecasts.Source: Forecast Results · pp. 6–7

Error sources

• Converting 1x1-degree grid predictions into latitude and longitude coordinates can add up to 50 km of forecast error.The paper reports that this margin remains competitive with NHC methods and proposes smaller grids or a neural conversion model to reduce it.Source: Forecast Results · p. 6, Limitations and Further Enhancement · p. 7

• 0.01 MSE and 0.11 RMSE were obtained on both the training and testing sets.Source: Conclusion · p. 7
• The model predicts the next hurricane location at 6-hour intervals using a recurrent neural network over a grid system.The interval matches the frequency of hurricane points in the Unisys Weather data.Source: Conclusion · p. 7
• The grid-based RNN can be trained to predict hurricanes of any type, unlike the previous sparse RNN, which does not work for looping hurricanes.The cited comparison notes that the evaluated storms were monotonic, although hurricane trajectories are not always monotonic.Source: Conclusion · p. 7
• Combining direction and distance features with a refined grid is reported to support high accuracy against currently employed NHC methods.The refined grid is described as reducing truncation errors and representing small-scale features more accurately.Source: Conclusion · p. 7
• The grid-based RNN is presented as a lightweight practical alternative to complex statistical-dynamical hurricane forecasting methods.The paper states that NHC methods can take hours to make a single prediction and identifies reduced grid-to-coordinate conversion as future work.Source: Conclusion · p. 7