A real time prediction methodology for hurricane evolution using LSTM recurrent neural networks

Abstract

Fast and accurate prediction of hurricane evolution from genesis onwards is needed to reduce loss of life and enhance community resilience. In this work, a novel model development methodology for predicting storm trajectory is proposed based on two classes of Recurrent Neural Networks (RNNs). The RNN models are trained on input features available in or derived from the HURDAT2 North Atlantic hurricane database maintained by the National Hurricane Center (NHC). The models use probabilities of storms passing through any location, computed from historical data. A detailed analysis of model forecasting error shows that Many-To-One prediction models are less accurate than Many-To-Many models owing to compounded error accumulation, with the exception of 6-hr predictions, for which the two types of model perform comparably. Application to 75 or more test storms in the North Atlantic basin showed that, for short-term forecasting up to 12 h, the Many-to-Many RNN storm trajectory prediction models presented herein are significantly faster than ensemble models used by the NHC, while leading to errors of comparable magnitude.

Base RNN model: LSTM

Recurrent Neural Networks (RNN) were innovated to extract pattern and context from sequences. RNNs are applied in a wide range of sequence related problems, including modeling and prediction of languages and sentiment, video tagging, a sequence prediction in time. The Long Short-Term Memory (LSTM), an RNN algorithm, was prescribed in [11] to tackle the vanishing gradient problem. Over long sequences, relevant past information may get lost or, equivalently, gradients may vanish while training a model using backpropagation. In an LSTM unit, past information may be retained via a cell state that passes through all LSTM layers. LSTM is the base RNN architecture of the models developed in the present work (Fig. 11).

Fig. 11

A representative LSTM unit; the unit belongs to the $q\mathrm{th}$ layer of the model. Corresponding input time step is l, i.e., at the input layer ($q=1$), $\vec {h}_l^0 = \vec {I}_l$ input features at the lth timestep. The symbols $\times$ and $+$ represent pointwise operation

An LSTM unit/ cell is shown in Fig. 11. The superscript in the vector variables indicates the layer number (this unit belongs to layer q of the model); the subscript denotes the corresponding input step l. A cell comprises four main components, the cell state (passing through the units, colored red), the forget gate (colored orange), the input gate (colored green) and the output gate (colored blue). The three gates basically apply the three activation functions (in the schematic, $\sigma$ and $\tanh$ represent sigmoid and hyperbolic tangent activation functions, respectively), each of which has a specific role in information propagation through the model. The cell state ($\overrightarrow{C^{(q)}_l}$) is the unique component of an LSTM RNN. The cell state passes through all timesteps $l=1, 2, ...$ of a given layer, and is therefore able to preserve information from the past and also accumulate new information with increasing l [14]. The cell shown in the diagram receives an input from the previous layer belonging to the same time step $\overrightarrow{h^{(q-1)}_l}$, and also from the previous time step in the same layer, $\overrightarrow{h^{(q)}_{l-1}}$. Based on these inputs to the cell, the forget gate dictates the part of the cell state to be discarded at the current unit. On the basis of these same inputs, input gate dictates the information from the present inputs to be added (marked by $+$) to the cell state. Consequently, after these operations, the cell state gets modified in the current LSTM unit ($\overrightarrow{C^{(q)}_{l-1}}\rightarrow \overrightarrow{C^{(q)}_l}$), which is the cell state received by the LSTM cell to the right, i.e., the next timestep in the same layer. The updated cell state also participates in obtaining the output from the current cell ($\overrightarrow{h^{(q)}_l}$) after the sigmoid activation is applied at the output gate.