Typhoon intensity prediction method and system based on quantum recurrent neural network

Abstract

The application provides a typhoon intensity prediction method and system based on a quantum recurrent neural network, which combines a quantum recurrent neural network and a single-head self-attention mechanism, the single-head self-attention mechanism can capture the correlation between different positions in a sequence, the quantum recurrent neural network can better process long-term time series data, and the combination of the single-head self-attention mechanism and the quantum recurrent neural network can fully utilize the advantages of the attention mechanism in capturing long-term dependence relationships and global information interactions in a sequence, thereby improving the prediction accuracy of typhoon intensity.

Description

Technical Field

[0001] This invention belongs to the technical field of typhoon forecasting, and particularly relates to a method and system for predicting typhoon intensity based on quantum recurrent neural networks. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Traditional typhoon intensity prediction methods are typically based on physical models and statistical methods, but these methods still face challenges in terms of accuracy and timeliness, such as the complexity of physical models, incomplete data, and limited extraction capabilities. With the rapid development of deep learning technology, more and more researchers are exploring its application in typhoon intensity prediction. Deep learning possesses powerful feature learning and pattern recognition capabilities; in typhoon intensity prediction, it can automatically extract features and build complex nonlinear models by learning from large amounts of data. Recurrent Neural Networks (RNNs) are suitable for processing sequential data and can capture the changing trends of typhoon paths and intensities over time, demonstrating good performance in short- and medium-term predictions.

[0004] However, deep learning requires a large amount of high-quality labeled data for training. However, it may be difficult to obtain accurate typhoon path and intensity data. Furthermore, the training and inference of deep learning models usually require a lot of computing resources. For real-time typhoon intensity prediction, the speed and efficiency of the model are a challenge. Summary of the Invention

[0005] To overcome the shortcomings of the prior art, this invention provides a typhoon intensity prediction method and system based on quantum recurrent neural networks. By combining quantum recurrent neural networks with a single-head self-attention mechanism, the advantages of the self-attention mechanism in capturing the correlation between different positions in the sequence can be fully utilized to improve the accuracy of typhoon intensity prediction and also improve the computation speed of the model.

[0006] To achieve the above objectives, a first aspect of the present invention provides a typhoon intensity prediction method based on a quantum recurrent neural network, comprising:

[0007] Obtain real-time relevant indicator data for the typhoon;

[0008] The real-time relevant index data of the typhoon is input into a trained quantum self-attention recurrent neural network to obtain the predicted typhoon intensity. The trained quantum self-attention recurrent neural network is obtained by training an initialized quantum self-attention recurrent neural network, which is composed of a first quantum path, a second quantum path, and a third quantum path in sequence. The first quantum path is a quantum single-head self-attention module used to capture the correlation of real-time typhoon-related index data; the second quantum path is a quantum recurrent neural network used to capture the connection between typhoon trajectories at different times; and the third quantum path is a quantum fully connected neural network used to integrate the first and second quantum paths, correlate the information from the quantum single-head self-attention module and the quantum recurrent neural network, and perform measurement and prediction.

[0009] A second aspect of the present invention provides a typhoon intensity prediction system based on a quantum recurrent neural network, comprising:

[0010] Acquisition module: Acquires real-time relevant indicator data for the typhoon;

[0011] Prediction Module: Inputs the real-time relevant index data of the typhoon into a trained quantum self-attention recurrent neural network to obtain the predicted typhoon intensity. The trained quantum self-attention recurrent neural network is obtained by training an initialized quantum self-attention recurrent neural network, which is composed of a first quantum path, a second quantum path, and a third quantum path in sequence. The first quantum path is a quantum single-head self-attention module used to capture the correlation of real-time typhoon-related index data; the second quantum path is a quantum recurrent neural network used to capture the connection between typhoon trajectories at different times; and the third quantum path is a quantum fully connected neural network used to integrate the first and second quantum paths, correlate the information from the quantum single-head self-attention module and the quantum recurrent neural network, and perform measurement and prediction.

[0012] A third aspect of the present invention provides a computer device, comprising: a processor, a memory, and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the computer device is running, the processor communicates with the memory via the bus, and when the machine-readable instructions are executed by the processor, a typhoon intensity prediction method based on a quantum recurrent neural network is executed.

[0013] A fourth aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, performs a typhoon intensity prediction method based on a quantum recurrent neural network.

[0014] The above one or more technical solutions have the following beneficial effects:

[0015] In this invention, by combining a quantum recurrent neural network and a single-head self-attention mechanism, the single-head self-attention mechanism can capture the correlation between different positions in the sequence, while the quantum recurrent neural network can better handle long-term time series data. The combination of the single-head self-attention mechanism and the quantum recurrent neural network can make full use of the advantages of the attention mechanism in capturing long-term dependencies and global information interaction in the sequence, thereby improving the accuracy of typhoon intensity prediction and increasing the computation speed of the model.

[0016] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0017] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0018] Figure 1 This is a flowchart of a typhoon intensity prediction method based on a quantum recurrent neural network according to Embodiment 1 of the present invention;

[0019] Figure 2 This is the data encoding method in a typhoon intensity prediction method based on a quantum recurrent neural network in Embodiment 1 of the present invention;

[0020] Figure 3 This is a structural diagram of the QRANN model in Embodiment 1 of the present invention;

[0021] Figure 4 This is a circuit diagram of the quantum recurrent neural network (QRNN) in Embodiment 1 of the present invention;

[0022] Figure 5 This is a circuit diagram of the quantum attention mechanism QSAM in Embodiment 1 of the present invention;

[0023] Figure 6 This is a circuit diagram of the quantum fully connected layer QFCL in Embodiment 1 of the present invention;

[0024] Figure 7 This is a flowchart of the model training process for the typhoon intensity prediction method based on a quantum recurrent neural network in Embodiment 1 of the present invention.

[0025] Figure 8 This is a comparison chart of the QRANN prediction results and the classical neural network prediction results in the typhoon intensity prediction method based on quantum recurrent neural networks in Embodiment 1 of the present invention. Detailed Implementation

[0026] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0027] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0028] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0029] Example 1

[0030] This embodiment discloses a typhoon intensity prediction method based on a quantum recurrent neural network, including:

[0031] Obtain real-time relevant indicator data for the typhoon;

[0032] The real-time relevant index data of the typhoon is input into a trained quantum self-attention recurrent neural network to obtain the predicted typhoon intensity. The trained quantum self-attention recurrent neural network is obtained by training an initialized quantum self-attention recurrent neural network, which is composed of a first quantum path, a second quantum path, and a third quantum path in sequence. The first quantum path is a quantum single-head self-attention module used to capture the correlation of real-time typhoon-related index data; the second quantum path is a quantum recurrent neural network used to capture the connection between typhoon trajectories at different times; and the third quantum path is a quantum fully connected neural network used to integrate the first and second quantum paths, correlate the information from the quantum single-head self-attention module and the quantum recurrent neural network, and perform measurement and prediction.

[0033] In recent years, thanks to more powerful computing devices, larger amounts of data, and technologies capable of training deeper networks, deep models have demonstrated superior performance compared to traditional methods in many fields, making deep learning a key technology in artificial intelligence. However, the further development of deep learning faces the following challenges: First, AI research based on deep neural networks and large-scale datasets places enormous demands on computational resources. With Moore's Law gradually losing its effectiveness, classical computing power is showing signs of fatigue. Simultaneously, in solving and optimizing deep models, the batch optimization strategy adopted due to computing power limitations struggles to find the global minimum. Furthermore, real-world data may contain patterns that classical algorithms and data mining mechanisms cannot easily detect, leading to a gap between the performance of classical deep learning algorithms and the expected optimal results. In particular, typhoons are violent air-sea interaction systems with probabilistic properties and a degree of uncertainty in their occurrence, making it extremely difficult to learn the hidden patterns of change within them.

[0034] Quantum computing processes information based on fundamental principles of uncertainty and nonlocality revealed by quantum mechanics, offering exponential speedup compared to classical computing for certain problems. In particular, the emergence of a series of quantum linear algebra algorithms with exponential speedup, including Fourier transforms, eigenvalue and eigenvector solving, and linear equation solving, provides powerful new tools for data analysis and deep learning. We believe that the quantum acceleration effect can significantly improve the performance of deep learning algorithms, especially when facing massive amounts of data. The combination of quantum computing and deep learning has the potential to reduce the cost of deep learning and provide new inspiration for deep learning algorithm research, ultimately leading to a revolution in artificial intelligence.

[0035] Quantum recurrent neural networks (QRNNs) have an exponentially dimensional feature space (2^32). n The QRNN's ability to represent complex time-series evolution patterns is unmatched by classical models. Therefore, QRNN has a stronger ability to learn hidden change patterns in complex time-series evolution systems. Applying QRNN to typhoon intensity forecasting is a feasible path to highlight the superiority of the QRNN model and explore its recent advantages.

[0036] The inclusion of quantum self-attention mechanism (QSAM) fully leverages the unique correlation of quantum entanglement in quantum circuits, enabling information transfer and sharing. In the quantum single-head self-attention mechanism model, quantum entanglement can be used to establish correlations between input sequences, allowing the model to better capture long-distance dependencies and contextual information within the sequences.

[0037] This embodiment provides a typhoon intensity forecasting method based on a quantum recurrent self-attention mechanism, enabling the prediction of typhoon intensity using a quantum computer-based deep learning algorithm. This promotes the practical application of the technology, leverages the advantages of quantum deep learning, and utilizes the superior learning ability of QRNN for complex change patterns hidden in time-series evolution systems. Combined with the quantum self-attention mechanism, this allows for the application of quantum deep learning in more real-world problems.

[0038] The core of this method lies in building a quantum algorithm. The implementation of quantum algorithms relies on fault-tolerant quantum computing built upon a large number of qubits. In other words, they are unsuitable for current first-generation quantum computers that only support a few hundred or a thousand qubits and a small number of noisy quantum gates. Therefore, developing a new quantum algorithm that can both leverage current advancements in quantum hardware and solve real-world problems becomes urgent.

[0039] Variational quantum algorithms emerged to address this need. At their core is an ansatz, which can be thought of as a template: it defines the qubits and the quantum gates that operate on them. Variational quantum circuits are constructed by repeatedly placing ansatz. The term "variational" comes from the quantum gates in ansatz that contain trainable parameters, such as the RX, RY, and RZ gates.

[0040]

[0041]

[0042]

[0043] Where I, X, Y, and Z represent the identity matrix, Pauli-X matrix, Pauli-Y matrix, and Pauli-Z matrix, respectively, and θ is a quantum parameter.

[0044] A parameterized quantum circuit is a quantum circuit formed by arranging single-qubit rotation gates and two-qubit gates in a certain layer according to a specific pattern and then superimposing them. Parameterization here refers to the fact that the single-qubit rotation gate contains trainable parameters. These parameters determine the influence of the quantum gate on the quantum state, such as phase changes, which in turn determine the measurement result of the quantum state.

[0045] In parameterized quantum circuits, the role of a parameterized quantum circuit in variational quantum algorithms is analogous to that of a neural network in deep learning; the parameters in the quantum circuit correspond to the adjustable weights in the neural network. More specifically, by measuring the expected values ​​of various observations on the parameterized quantum circuit, a corresponding loss function suitable for a specific problem can be constructed using these expected values. Using classical optimizer algorithms, the values ​​of the circuit parameters can be further adjusted based on the loss function values ​​corresponding to different circuit parameters. The new parameters are returned to the quantum circuit for further evaluation, and the parameters are updated again, continuing this cycle to improve the parameters and minimize the loss function corresponding to the problem, thus achieving an approximate solution for that type of problem.

[0046] Variable quantum circuits are similar to neural networks in some respects. First, by selecting an appropriate encoding strategy to encode the data, we can understand the input encoding as a nonlinear layer in a single-hidden-layer neural network. Second, the hierarchical structure built by the trainable Ansatz through repeated stacking can be interpreted as a deep linear neural network. Furthermore, the final measurement in the variable quantum circuit also introduces nonlinearity into the model. Therefore, the variable quantum model can also be called a quantum neural network model.

[0047] In the field of deep learning, neural networks with recurrent connections are called Recurrent Neural Networks (RNNs), a widely used neural network model in sequence data processing. Compared to traditional feedforward neural networks, RNNs have memory capabilities, enabling them to process time-dependent sequence data, model sequence data for sequence recognition and prediction, store information using iterative loops, effectively capture contextual information, and achieve transient dependency learning.

[0048] Recurrent Neural Networks (RNNs) are an extension of traditional feedforward neural networks, capable of handling variable-length input sequences. The core idea of ​​RNNs is to introduce recurrent connections, allowing the network to retain information from previous sequences while processing each element. It learns the hidden representation of the variable-length input sequence through internal recurrent hidden variables, where the output of the activation function at each time step depends on the output of the previous time step's hidden variable activation function. Given an input sequence x = (x1, x2, ..., x...), ... T Then, the cyclic update formula for the hidden variables of the RNN is as follows:

[0049] h tt =g(Wx t +Uh t-1 )

[0050] Where g is an activation function, W is the weight matrix of the latent variable input to this time step, and U is the weight matrix of the latent variable from the previous time step to the latent variable at this time step.

[0051] The core characteristic of recurrent neural networks (RNNs) is that the predicted value at a given time step is determined by both the current input data and information retained from past data. The purpose of training an RNN is to learn the optimal parameterized matrices U and W from the data. So how can we represent a classical RNN using a quantum recurrent structure? Or, in other words, how can we endow a quantum RNN with memory functionality? In the quantum circuit of this embodiment, the qubits are divided into two groups: the first group of qubits, denoted as n_1, represents the first half of the circuit, and the second group of qubits, denoted as n_2, represents the second half. At each time step, the first group of qubits is reinitialized and loaded with the current input data; the second group of qubits remains unchanged and is used to store information from past time steps, acting as a quantum memory. Specific details of the quantum structure construction are detailed in the implementation steps below.

[0052] The core of the attention mechanism lies in calculating a set of attention weights, which indicate the relative importance of different parts of the input. These weights are then used to calculate a weighted sum of the inputs, which is then passed to the rest of the model. Single-head self-attention is a variant of self-attention that uses only one attention head at each position to calculate the correlation between different positions in the sequence.

[0053] In traditional self-attention mechanisms, multiple attention heads are used to learn different semantic relationships. Each attention head computes an attention weight matrix, assigning a weight value to each position in the sequence, representing the relevance of that position to other positions. These multiple attention heads can be computed in parallel, thereby improving the model's efficiency and expressive power.

[0054] Single-head self-attention simplifies multi-head attention into a single mechanism, primarily used to model the relationships between different positions in a sequence. It associates each position with others by calculating attention weights. Specifically, single-head self-attention generates query, key, and value vectors by linearly transforming each position in the input sequence. Then, it calculates the click between the query and key vectors and performs a softmax operation to obtain attention weights. Finally, it weights and sums the attention weights with the value vectors to obtain the representation vector for each position.

[0055] Q = XW Q K = XW K V = XW V

[0056]

[0057] Single-head self-attention mechanisms are relatively simple to compute and have low computational complexity, making them suitable for handling large-scale data. Secondly, because there is only one attention head, the number of model parameters is smaller, reducing computational and storage costs. Furthermore, single-head self-attention mechanisms offer good interpretability, allowing visualization of the attention weights at each position, facilitating understanding of the model's decision-making process.

[0058] The quantum single-head self-attention mechanism aims to quantize the structure of the classical single-head self-attention mechanism in order to realize the function of the single-head self-attention mechanism. The specific construction method is described in the following technical solution steps.

[0059] like Figure 1As shown, the solution in this embodiment mainly consists of two parts: obtaining a segment of the path and intensity information of the current typhoon, preprocessing it, and then inputting it into the quantum recurrent self-attention neural network (QRANN) built in this embodiment to train the network and obtain the prediction result.

[0060] The specific technical solution of this implementation includes the following steps:

[0061] Step 1: Obtain the optimal path and intensity dataset of the typhoon, including six indicators: longitude, latitude, wind speed, air pressure, movement speed, and typhoon level at the current moment. These are labeled as follows: Where t represents time.

[0062] Step 2: Preprocess the acquired typhoon data, i.e., the training samples.

[0063] Step 3: Data preprocessing. Typhoon data is obtained and standardized. In this step, each indicator is standardized individually, ultimately yielding the data input into the quantum neural network.

[0064] The specific content of data preprocessing is as follows:

[0065]

[0066] Where S is the standard deviation of the sample, and μ is the mean of the sample.

[0067] Step 4: Establish a quantum cyclic self-attention mechanism neural network model, train the model, and use it for typhoon intensity prediction.

[0068] The construction of a quantum cyclic self-attention mechanism neural network. It mainly consists of a data encoding part, a parameterized quantum circuit part, and a quantum measurement part.

[0069] (1) Data encoding

[0070] like Figure 2 As shown, a quantum rotation gate is introduced to encode data into the angle of the quantum rotation gate. In this part, a single-qubit rotation gate, the RY gate, is used, and the data encoding method is as follows:

[0071]

[0072] Among them, X t This is the typhoon data captured at the current moment.

[0073] (2) Design of parameterized quantum circuit Ansatz

[0074] The encoded classical data evolves into a quantum state, which then passes through parameterized quantum circuits (PQCs), also known as Ansatz. PQCs are considered a new neural network paradigm in quantum computing. Unlike the weight matrix of classical neural networks, PQCs use the phase θ in the quantum gates as training parameters to evolve the encoded quantum state to the desired final quantum state. PQCs are unitary, and these unitary operations consist of parameterized quantum gates and two-qubit CNOT gates. The main function of the CNOT gates is to entangle the two involved qubits, thus generating multi-qubit entanglement.

[0075] Figure 3 The structure diagram of QRANN is shown, which mainly consists of three parts: the quantum recurrent neural network (QRNN), the quantum self-attention mechanism (QSAM), and the quantum fully connected layer (QFCL), with a total of 81 quantum parameters.

[0076] Figure 4 The diagram shows the QRNN quantum circuit structure, which uses 6 qubits and contains 30 quantum learning parameters. At each time step, 42 quantum gates are used: 30 single-qubit gates and 12 CNOT gates. The first 3 qubits are defined as the first group of qubits, used for data encoding. At each time step, the initial state of the first group of qubits is |0>, and real-time data is encoded into the quantum state through data encoding. The last 3 qubits are defined as the second group of qubits, used to carry information from the previous time step, and do not perform any operation.

[0077] This quantum circuit is based on the hardware-efficient quantum circuit proposed by Google in 2018. It can generate sufficient entanglement by using a quantum circuit composed of polynomial random selection of two-qubit gates. The two-qubit gates are multi-selected controlled quantum rotation gates and CNOT gates. Through the entanglement between qubits, its output state approximates the entire Hilbert space.

[0078] First, RX, RZ, and RX single-qubit rotation gates are placed sequentially on the six qubits to transform the input information in Hilbert space. The RX gate, acting on a single qubit, rotates it from the |0> state to the |1> state, and the rotation angle can be controlled. The RZ gate, acting on a single qubit, does not change the qubit's state but introduces a phase factor; changing the magnitude of this phase factor allows for phase adjustment of the quantum state.

[0079] Then, a combination of CNOT gates and RZ gates is applied to two adjacent qubits to achieve entanglement between the two qubits and enable information exchange.

[0080] This circuit can simultaneously achieve both quantum circuit representation and entanglement, meaning it possesses high entanglement and good expressiveness.

[0081] Figure 5 The diagram shows the QSAM quantum circuit structure, which uses 3 qubits and contains 15 quantum learning parameters. At each time step, 21 quantum gates are used: 15 single-qubit gates and 6 CNOT gates.

[0082] The construction of QSAM begins by introducing an RY gate into each of the three qubits. The RY gate acts on a single qubit, rotating it from the |0> state to the |1> state, and the rotation angle can be controlled. Because of the introduction of the complex number i, the RY gate can also realize the superposition of the qubit and the phase rotation of the quantum state. By applying the RY gate, Q, K, and V are obtained in the three qubits respectively. Then, CNOT gates and single-qubit gates are applied between pairs of qubits to perform information exchange and entanglement. The purpose is to train the three vectors Q, K, and V to be fully entangled, thereby achieving... This achieves the purpose of a single-head self-attention mechanism.

[0083] Figure 6 The diagram shows the QFCL quantum circuit structure, which uses 9 qubits and contains 36 quantum parameters. At each time step, 36 quantum gates are used: 18 single-qubit gates and 18 two-qubit controlled rotation gates.

[0084] The introduction of a quantum fully connected layer combines a quantum recurrent neural network with a quantum single-head self-attention mechanism, allowing the information learned by the quantum recurrent neural network to interact with the information learned by the quantum single-head self-attention mechanism. This fully leverages the advantages of both quantum recurrent neural networks and quantum single-head self-attention mechanisms to achieve accurate predictions.

[0085] The construction of a quantum fully connected layer follows the principle of fully entangled interaction between each qubit. Controlled RY gates are applied to each pair of qubits to enable information exchange between any two qubits, thus achieving the quantization of a classical fully connected layer. However, compared to classical fully connected layers, the construction of a quantum fully connected layer requires far fewer parameters, significantly reducing resource consumption compared to the thousands or even tens of thousands of parameters required for classical fully connected layers.

[0086] It is important to note that the angle of the single-qubit rotation gate is not fixed beforehand, but is updated during the iterative optimization process based on gradient descent. Furthermore, each PQC can be placed repeatedly to increase the depth of the neural network layers, thereby increasing the number of quantum parameters in the circuit. In this embodiment, the quantum circuit depth is set to 1.

[0087] (3) Quantum measurement

[0088] In the quantum measurement section, the probability of a single qubit collapsing to 1 is measured, and the measurement result is transformed into a predicted value through the inverse operations of data denormalization and normalization.

[0089] Let p be the probability that the qubit collapses to 1, and the measurement result of p can be expressed as:

[0090]

[0091] Here, ψ represents U(θ)|0>, which is the quantum state of the evolved parameterized quantum circuit acting on |0>. σ z The Pauli-Z matrix is... These are the measurement results for the quantum circuit.

[0092] This concludes the construction of the quantum neural network.

[0093] Figure 7 The diagram shows the training flowchart of a quantum neural network. Predictions are obtained through quantum measurement, and these are compared with the true values ​​to construct a loss function. In this embodiment, the chosen loss function is the mean squared error function, and its calculation formula is as follows:

[0094]

[0095] in, y is the predicted value, y is the standard value, and n is the number of samples.

[0096] By calculating the loss function, the loss value is passed to the optimizer, which optimizes the parameters according to the loss function. In this embodiment, the gradient descent method based on difference is selected to learn and optimize the model parameters.

[0097] It should be noted that for this gradient descent method, if there are N parameters in the quantum circuit, the quantum circuit needs to be run 2N times.

[0098] Predicting typhoon intensity is a challenging task, requiring accurate modeling and forecasting of time-series data. Combining a single-head self-attention mechanism with a recurrent neural network offers the following advantages:

[0099] (1) Capturing long-term dependencies: Recurrent neural networks may suffer from gradient vanishing or gradient exploding when dealing with long-term dependencies. However, the single-head self-attention mechanism can capture the correlation between different positions in the sequence, including long-term temporal dependencies. By introducing the self-attention mechanism, recurrent neural networks can better handle long-term time series data and improve their ability to model the evolution of typhoon intensity.

[0100] (2) Global Information Interaction: Recurrent neural networks can only transmit information through hidden states at each time step, while single-head self-attention mechanisms can perform global information interaction across the entire sequence at each time step. This means that the model can utilize information from all time steps in the sequence to predict the typhoon intensity at the current time step. Through the self-attention mechanism, recurrent neural networks can more comprehensively perceive and utilize the contextual information in the sequence, improving the accuracy of prediction.

[0101] (3) Parallel computation and parameter reduction: The computation of recurrent neural networks is performed step by step, and each time step depends on the hidden state of the previous time step. However, the single-head self-attention mechanism can compute the correlation between different positions in the sequence in parallel, thereby improving the computational efficiency of the model. In addition, since there is only one attention head, the number of model parameters is smaller, reducing computational and storage costs.

[0102] (4) Flexibility and interpretability: Combining the single-head self-attention mechanism with a recurrent neural network allows for flexible model design and adjustment. The weight allocation strategy of the attention mechanism, as well as the structure and hyperparameters of the recurrent neural network, can be adjusted according to the specific task requirements. In addition, the self-attention mechanism has good interpretability, allowing visualization of the attention weights at each location, which facilitates understanding the model's decision-making process and its contribution to typhoon intensity prediction.

[0103] In summary, combining a single-head self-attention mechanism with a recurrent neural network can fully leverage the advantages of attention mechanisms in capturing long-term dependencies and global information interaction in sequences, thereby improving the accuracy of typhoon intensity prediction. Furthermore, this combination offers advantages such as parallel computing, reduced parameters, flexibility, and interpretability, bringing better performance and application potential to typhoon intensity prediction tasks.

[0104] In this example, we selected Typhoon Haishen, a major typhoon of 2020, for verification. Typhoon Haishen was the first super typhoon of 2020. It formed in the Northwest Pacific Ocean on the evening of September 1 and made landfall on the coast of Gyeongsangnam-do, South Korea on September 7, with a maximum wind speed of 13 levels (40 m / s) at the center.

[0105] Since typhoon intensity is generally considered to be verified and compared based on the wind speed at the lower level, this embodiment mainly verifies and compares the wind speed at different stages of a typhoon.

[0106] This embodiment compares the typhoon intensity prediction models based on Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM). After multiple rounds of iterative training and optimization of the corresponding hyperparameters, the comparison between the model's prediction results and the actual observations is as follows. Figure 8 As shown, the predictions of CNN fluctuate significantly and the accuracy is not ideal. Compared with the CNN model, the LSTM model has better prediction stability and higher accuracy. The QRANN model built in this embodiment has even higher prediction accuracy and better stability.

[0107] Example 2

[0108] The purpose of this embodiment is to provide a typhoon intensity prediction system based on a quantum recurrent neural network, including:

[0109] Acquisition module: Acquires real-time relevant indicator data for the typhoon;

[0110] Prediction Module: Inputs the real-time relevant index data of the typhoon into a trained quantum self-attention recurrent neural network to obtain the predicted typhoon intensity. The trained quantum self-attention recurrent neural network is obtained by training an initialized quantum self-attention recurrent neural network, which is composed of a first quantum path, a second quantum path, and a third quantum path in sequence. The first quantum path is a quantum single-head self-attention module used to capture the correlation of real-time typhoon-related index data; the second quantum path is a quantum recurrent neural network used to capture the connection between typhoon trajectories at different times; and the third quantum path is a quantum fully connected neural network used to integrate the first and second quantum paths, correlate the information from the quantum single-head self-attention module and the quantum recurrent neural network, and perform measurement and prediction.

[0111] Example 3

[0112] The purpose of this embodiment is to provide a computing device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the above-described method.

[0113] Example 4

[0114] The purpose of this embodiment is to provide a computer-readable storage medium.

[0115] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the steps of the above-described method.

[0116] The steps and methods involved in the apparatuses of Embodiments 2, 3, and 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0117] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0118] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A typhoon intensity prediction method based on quantum recurrent neural networks, characterized in that, include: Obtain real-time relevant indicator data for the typhoon; The real-time relevant index data of the typhoon is input into the trained quantum self-attention recurrent neural network to obtain the prediction result of the typhoon intensity; wherein, the trained quantum self-attention recurrent neural network is obtained by training an initialized quantum self-attention recurrent neural network, which is composed of a first quantum route, a second quantum route and a third quantum route in sequence. The first quantum route is a quantum single-head self-attention module, which is used to capture the correlation of real-time typhoon-related index data. The second quantum route is a quantum recurrent neural network, used to capture the connection between typhoon trajectories at different times; the third quantum route is a quantum fully connected neural network, used to integrate the first and second quantum routes, linking the information of the quantum single-head self-attention module and the quantum recurrent neural network and performing measurement prediction. The real-time relevant data of the typhoon includes the typhoon's current longitude, latitude, wind speed, air pressure, movement speed, and typhoon level; The second quantum route includes a first set of qubits and a second set of qubits. The first set of qubits is used to encode real-time relevant index data of the typhoon into a quantum state. The second set of qubits is used to carry information from previous time steps. In the second quantum route, qubits are placed on each of the six qubits. RX , RZ , RX A single-qubit rotation gate transforms the input information in Hilbert space by applying a combination of CNOT and RZ gates to two adjacent qubits to achieve entanglement between them. In the first quantum route, an RY gate is introduced into each of the three qubits. By acting on the RY gate, the query vector, bond vector, and vector value are obtained in the three qubits respectively. The CNOT gate and the single-qubit gate are used to interact and entangle information between pairs of qubits, which are used to train the query vector, bond vector, and vector value. In the third quantum approach, controlled RY gates are applied to each pair of qubits, enabling information exchange between any two qubits.

2. The typhoon intensity prediction method based on quantum recurrent neural networks as described in claim 1, characterized in that, In the training of the quantum self-attention recurrent neural network, the output of the quantum self-attention recurrent neural network is compared with the true value to construct a mean square error function, and the mean square error function is used to train the quantum self-attention recurrent neural network.

3. A typhoon intensity prediction system based on quantum recurrent neural networks, characterized in that, include: Acquisition module: Acquires real-time relevant indicator data for the typhoon; Prediction module: Inputs the real-time relevant index data of the typhoon into the trained quantum self-attention recurrent neural network to obtain the prediction result of the typhoon intensity; wherein, the trained quantum self-attention recurrent neural network is obtained by training an initialized quantum self-attention recurrent neural network, which is composed of a first quantum path, a second quantum path and a third quantum path in sequence. The first quantum path is a quantum single-head self-attention module, which is used to capture the correlation of real-time typhoon-related index data; The second quantum route is a quantum recurrent neural network, used to capture the connection between typhoon trajectories at different times; the third quantum route is a quantum fully connected neural network, used to integrate the first and second quantum routes, linking the information of the quantum single-head self-attention module and the quantum recurrent neural network and performing measurement prediction. The real-time relevant data of the typhoon includes the typhoon's current longitude, latitude, wind speed, air pressure, movement speed, and typhoon level; The second quantum route includes a first set of qubits and a second set of qubits. The first set of qubits is used to encode real-time relevant index data of the typhoon into a quantum state. The second set of qubits is used to carry information from previous time steps. In the second quantum route, qubits are placed on each of the six qubits. RX , RZ , RX A single-qubit rotation gate transforms the input information in Hilbert space by applying a combination of CNOT and RZ gates to two adjacent qubits to achieve entanglement between them. In the first quantum route, an RY gate is introduced into each of the three qubits. By acting on the RY gate, the query vector, bond vector, and vector value are obtained in the three qubits respectively. The CNOT gate and the single-qubit gate are used to interact and entangle information between pairs of qubits, which are used to train the query vector, bond vector, and vector value. In the third quantum approach, controlled RY gates are applied to each pair of qubits, enabling information exchange between any two qubits.

4. A computer device, characterized in that, include: The computer device includes a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the computer device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, they perform the typhoon intensity prediction method based on a quantum recurrent neural network as described in any one of claims 1 to 2.

5. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, performs the typhoon intensity prediction method based on a quantum recurrent neural network as described in any one of claims 1 to 2.

Images