A multi-model integrated precipitation prediction method based on artificial intelligence

Abstract

The present invention relates to the field of electronic digital data processing, and in particular to a multi-mode integrated precipitation prediction method based on artificial intelligence. The method comprises collecting meteorological data, training a neural network based on chaos perturbation, training an LSTM based on quantum Monte Carlo, collecting new meteorological data, inputting new precipitation influencing factor features into the trained neural network based on chaos perturbation for feature fusion, inputting an actual feature vector set into the trained LSTM based on quantum Monte Carlo for prediction, and obtaining the precipitation on the target date. The existing multi-mode integrated precipitation prediction method has the problem of low prediction accuracy. The multi-mode integrated precipitation prediction method based on artificial intelligence provided by the present invention has high prediction accuracy.

Description

Technical Field

[0001] The present invention relates to the field of electronic digital data processing, and in particular to a multi-mode integrated precipitation prediction method based on artificial intelligence. Background Art

[0002] Studies have shown that among all kinds of natural disasters, meteorological disasters account for 55% of the deaths and 87% of the economic losses. Among them, precipitation-related meteorological disasters account for the largest proportion, such as drought, heavy rain, floods, mudslides, etc. It can be seen that precipitation is a very important meteorological element. Too much or too little will have a direct impact on production and life. Therefore, accurate precipitation forecasts in advance, estimation and inference of precipitation in a certain time and region in the future, can help prevent it, thereby reducing the impact of floods or droughts on people's production and life.

[0003] Traditional precipitation prediction methods mainly include: dynamics, statistics, and statistical-dynamic combination methods. People build precipitation prediction models for local areas based on these methods. However, due to the strong periodicity, correlation, and uncertainty of precipitation data, there are many difficulties in using traditional methods to predict precipitation. With the development of artificial intelligence technology, people began to use machine learning methods for prediction, such as random forest algorithms. Subsequently, with the further improvement of computing power and the development of deep learning technology, people began to use multi-mode integrated precipitation prediction methods for precipitation prediction. Multi-mode integrated precipitation prediction methods are precipitation prediction methods that consider multiple factors and use multiple algorithms when predicting.

[0004] The existing multi-mode integrated precipitation prediction methods mainly use neural networks to process meteorological data and use LSTM to predict the precipitation on the target date. However, when processing complex meteorological data, the existing neural network training methods are prone to problems such as gradient vanishing and gradient explosion. These problems greatly affect the training stability and prediction accuracy of the model. The traditional initialization method cannot effectively provide initial weights with sufficient nonlinear characteristics, resulting in a slow convergence speed during the training process and failure to fully explore the potential laws in meteorological data, which in turn affects the accuracy of precipitation prediction. In addition, the random initialization method is often used to train the existing LSTM, which may result in unreasonable initialization settings of LSTM parameters due to improper selection of random distribution and parameter settings, making the LSTM training unstable or even unable to proceed normally, ultimately affecting the prediction accuracy of LSTM.

[0005] Therefore, the existing multi-model integrated precipitation forecasting method has the problem of low prediction accuracy. Summary of the invention

[0006] The technical problem to be solved by the present invention is to overcome the shortcomings of the above-mentioned prior art and provide a multi-mode integrated precipitation prediction method based on artificial intelligence with high prediction accuracy.

[0007] In order to solve the above technical problems, the present invention provides a multi-mode integrated precipitation prediction method based on artificial intelligence, comprising:

[0008] S1. Collect meteorological data; the meteorological data includes precipitation influencing factor data and historical precipitation data; the historical precipitation data includes precipitation for multiple historical days; the precipitation influencing factor data is used as precipitation influencing factor features, and the precipitation for multiple historical days is used as precipitation features;

[0009] S2, inputting the precipitation influencing factor characteristics into a chaotic perturbation-based neural network and training the chaotic perturbation-based neural network, obtaining the trained chaotic perturbation-based neural network and the precipitation influencing factor characteristics after feature fusion after training, cascading the precipitation influencing factor characteristics after feature fusion with the precipitation characteristics to obtain a cascade feature set;

[0010] S3, inputting the cascade feature set into the LSTM based on quantum Monte Carlo and training the LSTM based on quantum Monte Carlo to obtain a trained LSTM based on quantum Monte Carlo;

[0011] S4. Collect new meteorological data, wherein the new meteorological data includes new precipitation influencing factor data and new historical precipitation data, wherein the new historical precipitation data includes new precipitation for multiple historical days, and use the new precipitation influencing factor data as new precipitation influencing factor features, and the new precipitation for multiple historical days as new precipitation features;

[0012] S5. Input the new precipitation influencing factor features into the trained chaotic perturbation-based neural network for feature fusion to obtain new precipitation influencing factor features after feature fusion, cascade the new precipitation influencing factor features after feature fusion with the new precipitation features to obtain an actual feature vector set, input the actual feature vector set into the trained quantum Monte Carlo-based LSTM for prediction to obtain the precipitation on the target date.

[0013] As a further improvement of the present invention: the training of the chaotic perturbation-based neural network in S2 includes:

[0014] S201, initializing the parameters of the neural network and the control parameters of the chaotic mapping;

[0015] S202. In each iteration, based on the The parameters of the iteration are the neural network The weight of the iteration The neural network Bias of iteration , the disturbance term is generated by the chaotic mapping function, and the calculation formula is:

[0016] ,

[0017] In the formula, is the disturbance term based on the parameters of the previous iteration, that is, the disturbance term of the parameters of the kth iteration; is the chaotic mapping function, is the amplitude factor of chaotic disturbance, Disturbance feedback coefficient, For the The perturbation term of the parameters of the iteration, is the period factor of the kth iteration;

[0018] S203. After each iteration, the loss function of the current neural network is calculated, and the fitness is calculated based on the loss function. The fitness measures the training performance of the current neural network. A high fitness value corresponds to a lower loss function. The calculation formula is:

[0019] ,

[0020] In the formula, is the fitness of the neural network at the kth iteration, is the loss function of the kth iteration of the neural network;

[0021] S204, adjusting the disturbance intensity according to the current fitness, a lower loss function fitness corresponds to a smaller disturbance intensity, and a larger loss function fitness corresponds to a larger disturbance intensity, so as to accelerate parameter adjustment. The adjustment method is expressed as:

[0022] ,

[0023] In the formula, is the disturbance term based on the parameters of the previous iteration, that is, the disturbance term of the parameters of the kth iteration; is the parameter update operation, is the disturbance term update factor, is the fitness of the neural network at the kth iteration, is the reverse adjustment factor for the kth iteration, For the neural network The fitness of the iteration;

[0024] S205, update the weights and biases of the neural network, and the update method is expressed as:

[0025] ,

[0026] In the formula, is the weight of the kth iteration of the neural network, The neural network The weight of the iteration, is the weight update amount of the kth iteration of the neural network, is the bias of the kth iteration of the neural network, The neural network The bias of the iteration, is the bias update amount of the kth iteration of the neural network;

[0027] S206. Dynamically adjust the update rate according to the change of the current fitness value. A higher fitness value will lead to a lower learning rate, and vice versa. The adjustment method is expressed as:

[0028] ,

[0029] In the formula, is the learning rate of the kth iteration of the neural network, is the initial learning rate of the neural network, is the adaptability factor, is the fitness of the neural network at the kth iteration;

[0030] S207, repeating iterations S202 to S206 until a preset condition for stopping iteration is met, which means that the training of the neural network based on chaotic perturbation is completed.

[0031] Preferably, the step S202 The calculation formula is:

[0032] ,

[0033] In the formula, is the amplitude factor of chaotic disturbance, For the The random variable of the first chaotic map of iterations, For the The random variable of the second chaotic map of iterations.

[0034] Preferably, the S203 The calculation formula is:

[0035] ,

[0036] In the formula, is the number of meteorological data samples input to the neural network in the current batch, is the true label of the i-th meteorological data sample; is the prediction output obtained by extracting features of the i-th meteorological data sample based on the chaotic perturbation neural network and classifying it with the preset Softmax function at the k-th iteration. is the weighting factor of the loss function, for norm, is the feature vector of the i-th sample, is the output feature value of the neural network for the i-th sample.

[0037] Preferably, the step S205 , The calculation formulas are:

[0038] ,

[0039] In the formula, is the update adjustment factor of the neural network; is the disturbance term based on the parameters of the previous iteration, that is, the disturbance term of the parameters of the kth iteration; is the feature vector of the sample of the kth iteration, is the learning rate of the kth iteration of the neural network.

[0040] As a further improvement of the present invention: the training of the quantum Monte Carlo-based LSTM in S3 includes:

[0041] S301, initialize the parameters of LSTM and quantum Monte Carlo algorithm;

[0042] S302, perform feature encoding of meteorological data, assuming that the meteorological data input to LSTM is a cascade feature set obtained by cascading the precipitation influencing factor features and precipitation features after feature fusion, defined as , through multi-layer nonlinear mapping, the feature-encoded meteorological data is obtained as follows:

[0043] ,

[0044] In the formula, is the feature vector of the tth time step, For multi-layer nonlinear mapping operations, is the weight matrix of multi-layer nonlinear mapping, is the eigenvector of the t-1th time step, is the state information of the previous time step;

[0045] S303, LSTM is trained through the back propagation algorithm to optimize the weight and bias of the network; LSTM effectively learns the long-term dependencies in meteorological data through its unique gating mechanism. During the training process, LSTM will continuously adjust the network parameters to minimize the prediction error. The loss function of LSTM is expressed as:

[0046] ,

[0047] In the formula, is the loss function of LSTM, is the weight coefficient of the first loss, is the number of training samples, is the predicted value of the i-th sample, is the true value of the i-th sample, is the weight coefficient of the second loss, Hamiltonian optimized for quantum;

[0048] S304. In order to make the feature weights adaptively adjusted according to the current environmental conditions at every moment, a strategy based on quantum optimization is adopted to find the optimal weight value by measuring the quantum state. The quantum optimized Hamiltonian The calculation formula is:

[0049] ,

[0050] In the formula, is the feature dimension; is the importance parameter of the feature for regression prediction based on quantum Monte Carlo LSTM, and is the training parameter; Indicates that a specific feature Below, the output of the regression prediction function;

[0051] S305. During the training process, the learning rate of LSTM is adaptively adjusted according to the feedback of the loss function of LSTM. LSTM based on quantum Monte Carlo can improve the convergence speed in the later stage of training and avoid premature convergence in the early stage. The adjustment method is expressed as:

[0052] ,

[0053] In the formula, is the learning rate of the LSTM at the k+1th iteration, is the learning rate of LSTM at the kth iteration, is the learning rate adjustment factor of LSTM, is the loss value of LSTM at the kth iteration, For LSTM The loss value of the iteration;

[0054] S306. Repeat S302 to S305 until the preset stop iteration condition is met, which means that the LSTM training based on quantum Monte Carlo is completed.

[0055] Preferably, the S303 The calculation formula is:

[0056] ,

[0057] In the formula, is the number of quantum sampling; is the quantum bit result in the quantum sampling process at the tth time step; is the activation function of LSTM, is the input feature of the i-th sample.

[0058] The beneficial effects of the present invention are as follows: The multi-mode integrated precipitation prediction method based on artificial intelligence provided by the present invention has high prediction accuracy.

[0059] This method uses a chaotic perturbation-based optimization neural network for feature fusion. When training the chaotic perturbation-based optimization neural network, a disturbance term is introduced through chaotic mapping to optimize the update of parameters during the neural network training process. The chaotic perturbation term can effectively avoid the common gradient vanishing, gradient explosion and local optimal solution problems in the traditional neural network training process, and can ensure the stability of the neural network training when processing complex meteorological data. At the same time, the initial weights and biases of the neural network are adjusted through chaotic perturbations, so that the initial parameters of the network have sufficient nonlinear characteristics, which helps to enhance the expressiveness of the network in the precipitation prediction task. Moreover, in each training iteration, the intensity of the disturbance term is dynamically adjusted. According to the feedback of the current loss function, the disturbance term is gradually increased, thereby accelerating the convergence of the neural network, reducing the convergence time caused by unstable training, and improving the training efficiency of the model, thereby improving the prediction accuracy of the method.

[0060] This method also uses a quantum Monte Carlo-based LSTM to predict the precipitation on the target date. When training the quantum Monte Carlo-based LSTM, the quantum Monte Carlo algorithm uses quantum sampling to improve the accuracy of LSTM network initialization, avoiding the problem of training difficulties caused by improper initialization in the existing LSTM, which leads to reduced prediction accuracy.

[0061] Therefore, the prediction accuracy of this method is higher. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 A step diagram of the method provided by the present invention. DETAILED DESCRIPTION

[0063] The specific implementation modes of the present invention are further described in detail below in conjunction with the accompanying drawings.

[0064] Various sequences in the present invention and the drawings, such as “S1”, are only for the purpose of clearly describing a certain embodiment and are not meant to be a necessary sequence unless otherwise specified that a certain sequence must be followed.

[0065] (1) Long Short-Term Memory (LSTM) is a special type of recurrent neural network (RNN) that can effectively solve the long-term dependency problem when processing time series data. LSTM solves the long-term dependency problem by introducing memory cells and three gating mechanisms. LSTM contains one or more memory cells for storing sequence information. Each cell has three important gating mechanisms: input gate, forget gate, and output gate. The three gating mechanisms work together to control the storage, update, or discard of information in the memory cell, allowing LSTM to selectively ignore unimportant information. The input gate controls the addition of new information and determines how much of the current input information should be written into the memory cell. It uses the Sigmoid activation function to determine which information should be updated and uses the Tanh activation function to generate new candidate memories. The forget gate determines how much information in the memory cell at the current moment should be forgotten. It receives the current input and the hidden state of the previous moment and outputs a value between 0 and 1. 1 means completely retained and 0 means completely forgotten. The output gate controls the content of the memory cell to be output to the next time step or as the output at the current moment. It will decide how much information to output based on the state of the memory unit and the current input; the memory unit is updated based on the operation of the forget gate and the input gate. The forget gate determines the old information that needs to be discarded, and the input gate determines the addition of new information.

[0066] like Figure 1 As shown, the present invention provides a multi-mode integrated precipitation prediction method based on artificial intelligence, including:

[0067] S1. Collect and store meteorological data; the meteorological data includes precipitation influencing factor data and historical precipitation data; the precipitation influencing factor data is used as precipitation influencing factor features, and the precipitation of multiple historical days is used as precipitation features;

[0068] Meteorological data are collected by weather stations, satellites, and weather radars.

[0069] The data on factors affecting precipitation include time characteristics and auxiliary meteorological factors. Time characteristics include time period and seasonal changes, and auxiliary meteorological factors include temperature, humidity, wind speed, and air pressure.

[0070] The historical precipitation data includes the precipitation for multiple historical days and the historical precipitation range, historical precipitation intensity, and historical precipitation speed corresponding to the precipitation for multiple historical days.

[0071] S2, inputting the precipitation influencing factor characteristics into a chaotic perturbation-based neural network and training the chaotic perturbation-based neural network, obtaining the trained chaotic perturbation-based neural network and the precipitation influencing factor characteristics after feature fusion after training, cascading the precipitation influencing factor characteristics after feature fusion with the precipitation characteristics to obtain a cascade feature set;

[0072] The neural network algorithm based on chaotic perturbation means that each parameter in the neural network will be updated through the perturbation term generated by the chaotic mapping. The perturbation term not only has nonlinear characteristics, but also can adaptively adjust the intensity of the perturbation according to the local behavior of the loss function, thereby accelerating the convergence process and improving the accuracy of meteorological data processing. The algorithm is inspired by the principle of chaotic systems. Chaotic mapping has unique advantages in optimization problems due to its unpredictability and nonlinear characteristics. In the process of neural network parameter optimization, it can avoid the troubles of local minima and saddle points, thereby effectively accelerating the training process of the neural network.

[0073] Training the chaotic perturbation-based neural network includes:

[0074] S201, initializing the parameters of the neural network and the control parameters of the chaotic map; in the initialization stage, the weight and bias parameters of the neural network are affected by the chaotic disturbance, and the control parameters of the chaotic map need to be initialized at the same time to ensure that the initial parameter distribution of the neural network has sufficient nonlinear characteristics. The parameter initialization method of the neural network is expressed as:

[0075] ,

[0076] In the formula, is the initial weight of the neural network; is the first initialization amplitude factor, ; is the first chaotic map random variable, is the second chaotic map random variable, is the initial bias of the neural network, is the second initialization amplitude factor, ;

[0077] The control parameters of the chaotic map include the first chaotic map random variable With the second chaotic map random variable , the control parameters of the chaotic map are initialized as follows:

[0078] ,

[0079] In the formula, is the first chaotic map random variable, is the first random variable constant, is the first randomly initialized angle variable, is the second chaotic map random variable, is the second random variable constant, ; Angle variable for the second random initialization;

[0080] S202, in each iteration process, based on the last (i.e. The parameters of the neural network are the The weight of the iteration The neural network Bias of iteration , the disturbance term is generated by the chaotic mapping function. The disturbance term not only reflects the nonlinear dynamic behavior of the current parameters, but also considers the historical influence of the system. The calculation formula is:

[0081] ,

[0082] In the formula, is the disturbance term based on the parameters of the previous iteration, that is, the disturbance term of the parameters of the kth iteration; is the chaotic mapping function, is the amplitude factor of the chaotic disturbance, set to =0.5; is the disturbance feedback coefficient, set to =0.6; For the The perturbation term of the parameters of the iteration, is the period factor of the kth iteration;

[0083] The chaos map function uses a chaos map to generate disturbance terms. The calculation formula is:

[0084] ,

[0085] In the formula, is the amplitude factor of chaotic disturbance, For the The random variable of the first chaotic map of iterations, For the The random variable of the second chaotic map of iterations;

[0086] The chaotic disturbance term can gradually increase with the iteration, thereby promoting the update of the parameters of the neural network. The calculation formula is:

[0087] ,

[0088] In the formula, The random variable of the second chaotic map at the kth iteration, is the update factor of the third chaotic perturbation term, For the The random variable of the second chaotic map of iterations, is the update factor of the first chaotic perturbation term, For the The perturbation term of the parameters of the iteration, is the first chaotic map random variable of the kth iteration; is the update factor of the fourth chaotic perturbation term, ; is the update factor of the second chaotic perturbation term, ;

[0089] The periodic factor adjusts the amplitude of the disturbance through an additional oscillation period, so that the disturbance term is not only affected by the current parameter state, but also by the feedback of historical disturbances, thereby increasing the nonlinear dynamics of the disturbance term. The calculation formula is:

[0090] ,

[0091] In the formula, is the period factor of the kth iteration; is the adjustment parameter of the period factor, set to =0.5; is the factor of change of the first control cycle, Set to 0.2; For the The random variable of the second chaotic map of iterations; is the factor of the second control cycle change, Set to 0.5; For the The random variable of the first chaotic map of iterations.

[0092] S203. After each iteration, the loss function of the current neural network is calculated, and the fitness is calculated based on the loss function. The fitness measures the training performance of the current neural network. A high fitness value corresponds to a lower loss function. The calculation formula is:

[0093] ,

[0094] In the formula, is the fitness of the neural network at the kth iteration; is the loss function of the kth iteration of the neural network;

[0095] The loss function of the neural network calculates the norm of the features of the neural network, thereby penalizing the difference between the input features and the output features, and enhancing the attention of the neural network based on chaotic perturbation to the feature space. The calculation formula is:

[0096] ,

[0097] In the formula, is the number of meteorological data samples input to the neural network in the current batch, is the true label of the i-th meteorological data sample; is the prediction output obtained by the preset Softmax function classification after the feature extraction of the i-th meteorological data sample by the chaotic perturbation neural network at the k-th iteration; is the weighting factor of the loss function, set to =0.3; for norm, is the characteristic vector of the i-th sample, including the data of factors affecting precipitation, is the output feature value of the neural network for the i-th sample;

[0098] S204, adjusting the disturbance intensity according to the current fitness, a lower loss function fitness corresponds to a smaller disturbance intensity, and a larger loss function fitness corresponds to a larger disturbance intensity, so as to accelerate parameter adjustment. The adjustment method is expressed as:

[0099] ,

[0100] In the formula, is the disturbance term based on the parameters of the previous iteration, that is, the disturbance term of the parameters of the kth iteration; It is the parameter update operation; is the disturbance term update factor, which is used to control the change of disturbance intensity and is set to =0.05; is the fitness of the neural network at the kth iteration; is the reverse adjustment factor for the kth iteration, set to =0.2; For the neural network The fitness of the iteration;

[0101] S205, update the weights and biases of the neural network, and the update method is expressed as:

[0102] ,

[0103] In the formula, is the weight of the kth iteration of the neural network, The neural network The weight of the iteration, is the weight update amount of the kth iteration of the neural network, is the bias of the kth iteration of the neural network, The neural network The bias of the iteration, is the bias update amount of the kth iteration of the neural network;

[0104] The weight update and bias update of the neural network are calculated by the perturbation term of the parameters. , The calculation formulas are:

[0105] ,

[0106] In the formula, is the update adjustment factor of the neural network, set to =0.05; is the disturbance term based on the parameters of the previous iteration, that is, the disturbance term of the parameters of the kth iteration; is the feature vector of the sample of the kth iteration, is the learning rate of the kth iteration of the neural network.

[0107] S206. Dynamically adjust the update rate according to the change of the current fitness value. A higher fitness value will lead to a lower learning rate, and vice versa. The adjustment method is expressed as:

[0108] ,

[0109] In the formula, is the learning rate of the kth iteration of the neural network; is the initial learning rate of the neural network, set to =0.01; is the adaptability factor, which determines the sensitivity of learning rate adjustment; is the fitness of the neural network at the kth iteration;

[0110] S207, repeating iterations S202 to S206 until a preset stop iteration condition is met, indicating that the neural network training based on chaotic perturbation is completed. The preset stop iteration condition is reaching a preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times.

[0111] S3, inputting the cascade feature set into the LSTM based on quantum Monte Carlo and training the LSTM based on quantum Monte Carlo to obtain a trained LSTM based on quantum Monte Carlo;

[0112] The use of cascaded feature sets can make the precipitation prediction process of LSTM based on quantum Monte Carlo affected by time characteristics and auxiliary meteorological factors, thereby improving the accuracy and reliability of precipitation prediction.

[0113] Quantum Monte Carlo-based LSTM refers to the optimization of LSTM using the Quantum Monte Carlo algorithm (QMC). Specifically, the Quantum Monte Carlo algorithm provides an efficient initialization method for LSTM training through quantum sampling.

[0114] The training of the quantum Monte Carlo-based LSTM in S3 includes:

[0115] S301, initialize the parameters of LSTM and quantum Monte Carlo algorithm; the initialization method is expressed as:

[0116] ,

[0117] In the formula, is the initialized LSTM hidden layer state vector, is the activation function of LSTM, is the weight matrix of LSTM, is the input feature vector at the time of initialization, specifically, the cascade feature set obtained by concatenating the precipitation influencing factor features after feature fusion with the precipitation features, such as the cascade feature vector obtained by concatenating the precipitation influencing factor features after feature fusion with the historical 30-day precipitation, and the label is the precipitation on the target date; is the bias term of LSTM;

[0118] Assume the number of quantum sampling in the quantum Monte Carlo algorithm is , the initialization method is expressed as:

[0119] ,

[0120] In the formula, is the state of quantum Monte Carlo sampling at initialization, is the tth sampling result.

[0121] S302, perform feature encoding of meteorological data, assuming that the meteorological data input to LSTM is a cascade feature set obtained by cascading the precipitation influencing factor features and precipitation features after feature fusion, defined as , through multi-layer nonlinear mapping, the feature-encoded meteorological data is obtained as follows:

[0122] ,

[0123] In the formula, is the feature vector of the tth time step, For multi-layer nonlinear mapping operations, is the weight matrix of multi-layer nonlinear mapping, is the eigenvector of the t-1th time step, is the state information of the previous time step;

[0124] The feature encoding method in the existing LSTM fails to effectively enhance the expressive power of input features, making it difficult to accurately reflect the interaction of various meteorological factors in precipitation prediction, thus affecting the prediction accuracy of the model. The LSTM based on quantum Monte Carlo performs feature encoding in S302, and cooperates with the cascade feature set obtained by concatenating the precipitation influencing factor features after feature fusion with the precipitation features, thereby enhancing the representation power of the input features. When dealing with the spatiotemporal changes in precipitation, it can effectively capture the key patterns of historical precipitation and strengthen the impact of these features.

[0125] S303, LSTM is trained through the back propagation algorithm to optimize the weight and bias of the network; LSTM effectively learns the long-term dependencies in meteorological data through its unique gating mechanism. During the training process, LSTM will continuously adjust the network parameters to minimize the prediction error. The loss function of LSTM is expressed as:

[0126] ,

[0127] In the formula, is the loss function of LSTM, is the weight coefficient of the first loss, Set to 0.3, is the number of training samples, is the predicted value of the i-th sample, is the true value of the i-th sample, is the weight coefficient of the second loss, Set to 0.7, is the quantum optimized Hamiltonian, which represents the optimization error of the current feature weight;

[0128] The predicted value of a sample depends not only on the output of LSTM, but also on the output calculated by the quantum Monte Carlo algorithm. The calculation formula is:

[0129] ,

[0130] In the formula, is the number of quantum sampling; is the quantum bit result in the quantum sampling process at the tth time step. Multiple samples are generated by random sampling using quantum bits. These samples are mapped to the classical data space through quantum measurement. For example, assuming that the result of quantum sampling is to sample certain features of precipitation data, the quantum system will extract key time series features based on the changing trend of precipitation and related meteorological factors (such as temperature and humidity). Through quantum Monte Carlo sampling, the LSTM based on quantum Monte Carlo can identify abnormal patterns of certain days in the precipitation of the past 30 days (such as continuous heavy rain days or drought periods) and strengthen the influence of these key features in the feature fusion process; is the activation function of LSTM, is the input feature of the i-th sample;

[0131] The quantum bit results in the quantum sampling process are optimized based on the quantum Monte Carlo algorithm, aiming to reduce the computational overhead of the quantum Monte Carlo-based LSTM through efficient sampling. The calculation formula is:

[0132] ,

[0133] In the formula, It is the basic function of quantum sampling, specifically the Hadamard gate function, which converts the quantum bit from the standard ground state to the superposition state and defines the generation rule of the quantum bit; is the adjustment coefficient of quantum sampling; is the quantum bit result in the quantum sampling process at the t-1th time step.

[0134] During the LSTM training process, the loss function is minimized by the gradient descent method, and the LSTM weight update method is expressed as:

[0135] ,

[0136] In the formula, is the LSTM weight at the k+1th iteration, is the LSTM weight at the kth iteration, is the learning rate of LSTM, is the gradient of the LSTM loss function with respect to the weights.

[0137] S304. In order to make the feature weights adaptively adjusted according to the current environmental conditions at every moment, a strategy based on quantum optimization is adopted to find the optimal weight value by measuring the quantum state. The specific goal is to minimize the quantum optimized Hamiltonian. The quantum optimized Hamiltonian The calculation formula is:

[0138] ,

[0139] In the formula, is the feature dimension; is the importance parameter of the feature for predicting the LSTM based on quantum Monte Carlo, and is the training parameter; Indicates that a specific feature Below, the output of the regression prediction function;

[0140] When training the existing LSTM, due to the high dimensionality and complexity of meteorological data, LSTM is prone to fall into local minima or saddle points, causing training stagnation and affecting the final prediction performance. Traditional optimization methods (such as gradient descent) are prone to local optimal solutions in such high-dimensional, nonlinear problems, and thus cannot fully optimize the performance of the model, reducing the accuracy of precipitation prediction. In S304, the quantum optimization process minimizes the quantum optimized Hamiltonian, allowing the LSTM model to converge quickly during the training process of high-dimensional meteorological data, avoiding the problem that traditional LSTM is prone to falling into local minima, thereby improving the accuracy of precipitation prediction.

[0141] S305. During the training process, the learning rate of LSTM is adaptively adjusted according to the feedback of the loss function of LSTM. LSTM based on quantum Monte Carlo can improve the convergence speed in the later stage of training and avoid premature convergence in the early stage. The adjustment method is expressed as:

[0142] ,

[0143] In the formula, is the learning rate of the LSTM at the k+1th iteration, is the learning rate of LSTM at the kth iteration, is the learning rate adjustment factor of LSTM, Set to 0.05; is the loss value of LSTM at the kth iteration, For LSTM The loss value of the iteration;

[0144] S306, repeating S302-S305 until the preset stop iteration condition is met, which means that the LSTM training based on quantum Monte Carlo is completed. The preset stop iteration condition is reaching the preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times.

[0145] After LSTM training is completed, in the last layer of LSTM, the state vector of the hidden layer (which contains the memory of the historical 30-day precipitation data) will be passed to a fully connected layer (which can be a simple linear layer or an activation function), which is used to convert the output of LSTM into a specific value, that is, the precipitation on the target date.

[0146] S4. Collect new meteorological data, wherein the new meteorological data includes new precipitation influencing factor data and new historical precipitation data, wherein the new historical precipitation data includes new precipitation for multiple historical days, and use the new precipitation influencing factor data as new precipitation influencing factor features, and the new precipitation for multiple historical days as new precipitation features;

[0147] S5. Input the new precipitation influencing factor features into the trained chaotic perturbation-based neural network for feature fusion to obtain new precipitation influencing factor features after feature fusion, cascade the new precipitation influencing factor features after feature fusion with the new precipitation features to obtain an actual feature vector set, input the actual feature vector set into the trained quantum Monte Carlo-based LSTM for prediction to obtain the precipitation on the target date.

Claims

1. A multi-mode integrated precipitation prediction method based on artificial intelligence, characterized in that: include: S1. Collect meteorological data; the meteorological data includes precipitation influencing factor data and historical precipitation data; the historical precipitation data includes precipitation for multiple historical days; the precipitation influencing factor data is used as precipitation influencing factor features, and the precipitation for multiple historical days is used as precipitation features; S2, inputting the precipitation influencing factor characteristics into a chaotic perturbation-based neural network and training the chaotic perturbation-based neural network, obtaining the trained chaotic perturbation-based neural network and the precipitation influencing factor characteristics after feature fusion after training, cascading the precipitation influencing factor characteristics after feature fusion with the precipitation characteristics to obtain a cascade feature set; S3, inputting the cascade feature set into the LSTM based on quantum Monte Carlo and training the LSTM based on quantum Monte Carlo to obtain a trained LSTM based on quantum Monte Carlo; S4. Collect new meteorological data, wherein the new meteorological data includes new precipitation influencing factor data and new historical precipitation data, wherein the new historical precipitation data includes new precipitation for multiple historical days, and use the new precipitation influencing factor data as new precipitation influencing factor features, and the new precipitation for multiple historical days as new precipitation features; S5. Input the new precipitation influencing factor features into the trained chaotic perturbation-based neural network for feature fusion to obtain new precipitation influencing factor features after feature fusion, cascade the new precipitation influencing factor features after feature fusion with the new precipitation features to obtain an actual feature vector set, input the actual feature vector set into the trained quantum Monte Carlo-based LSTM for prediction to obtain the precipitation on the target date.

2. The multi-mode integrated precipitation prediction method based on artificial intelligence according to claim 1 is characterized in that: The step of training the chaotic perturbation-based neural network in S2 includes: S201, initializing the parameters of the neural network and the control parameters of the chaotic mapping; S202. In each iteration, based on the The parameters of the iteration are the neural network The weight of the iteration The neural network Bias of iteration , the disturbance term is generated by the chaotic mapping function, and the calculation formula is: ， In the formula, is the disturbance term based on the parameters of the previous iteration, that is, the disturbance term of the parameters of the kth iteration; is the chaotic mapping function, is the amplitude factor of chaotic disturbance, is the disturbance feedback coefficient, For the The perturbation term of the parameters of the iteration, is the period factor of the kth iteration; S203. After each iteration, the loss function of the current neural network is calculated, and the fitness is calculated based on the loss function. The fitness measures the training performance of the current neural network. A high fitness value corresponds to a lower loss function. The calculation formula is: ， In the formula, is the fitness of the neural network at the kth iteration, is the loss function of the kth iteration of the neural network; S204, adjusting the disturbance intensity according to the current fitness, a lower loss function fitness corresponds to a smaller disturbance intensity, and a larger loss function fitness corresponds to a larger disturbance intensity, so as to accelerate parameter adjustment. The adjustment method is expressed as: ， In the formula, is the disturbance term based on the parameters of the previous iteration, that is, the disturbance term of the parameters of the kth iteration; is the parameter update operation, is the disturbance term update factor, is the fitness of the neural network at the kth iteration, is the reverse adjustment factor for the kth iteration, For the neural network The fitness of the iteration; S205, update the weights and biases of the neural network, and the update method is expressed as: ， In the formula, is the weight of the kth iteration of the neural network, The neural network The weight of the iteration, is the weight update amount of the kth iteration of the neural network, is the bias of the kth iteration of the neural network, The neural network The bias of the iteration, is the bias update amount of the kth iteration of the neural network; S206. Dynamically adjust the update rate according to the change of the current fitness value. A higher fitness value will lead to a lower learning rate, and vice versa. The adjustment method is expressed as: ， In the formula, is the learning rate of the kth iteration of the neural network, is the initial learning rate of the neural network, is the adaptability factor, is the fitness of the neural network at the kth iteration; S207, repeating iterations S202 to S206 until a preset condition for stopping iteration is met, which means that the training of the neural network based on chaotic perturbation is completed.

3. The multi-mode integrated precipitation prediction method based on artificial intelligence according to claim 2 is characterized in that: As described in S202 The calculation formula is: ， In the formula, is the amplitude factor of chaotic disturbance, For the The random variable of the first chaotic map of iterations, For the The random variable of the second chaotic map of iterations.

4. The multi-mode integrated precipitation prediction method based on artificial intelligence according to claim 2 is characterized in that: S203 The calculation formula is: ， In the formula, is the number of meteorological data samples input to the neural network in the current batch, is the true label of the i-th meteorological data sample; is the prediction output obtained by extracting features of the i-th meteorological data sample based on the chaotic perturbation neural network and classifying it with the preset Softmax function at the k-th iteration. is the weighting factor of the loss function, for norm, is the feature vector of the i-th sample, is the output feature value of the neural network for the i-th sample.

5. The multi-mode integrated precipitation prediction method based on artificial intelligence according to claim 2 is characterized in that: S205 , The calculation formulas are: ， In the formula, is the update adjustment factor of the neural network; is the disturbance term based on the parameters of the previous iteration, that is, the disturbance term of the parameters of the kth iteration; is the feature vector of the sample of the kth iteration, is the learning rate of the kth iteration of the neural network.

6. The multi-mode integrated precipitation prediction method based on artificial intelligence according to any one of claims 1 to 5, characterized in that: The training of the quantum Monte Carlo-based LSTM in S3 includes: S301, initialize the parameters of LSTM and quantum Monte Carlo algorithm; S302, perform feature encoding of meteorological data, assuming that the meteorological data input to LSTM is a cascade feature set obtained by cascading the precipitation influencing factor features and precipitation features after feature fusion, defined as , through multi-layer nonlinear mapping, the feature-encoded meteorological data is obtained as follows: ， In the formula, is the feature vector of the tth time step, For multi-layer nonlinear mapping operations, is the weight matrix of multi-layer nonlinear mapping, is the eigenvector of the t-1th time step, is the state information of the previous time step; S303, LSTM is trained through the back propagation algorithm to optimize the weight and bias of the network; LSTM effectively learns the long-term dependencies in meteorological data through its unique gating mechanism. During the training process, LSTM will continuously adjust the network parameters to minimize the prediction error. The loss function of LSTM is expressed as: ， In the formula, is the loss function of LSTM, is the weight coefficient of the first loss, is the number of training samples, is the predicted value of the i-th sample, is the true value of the i-th sample, is the weight coefficient of the second loss, Hamiltonian optimized for quantum; S304. In order to make the feature weights adaptively adjusted according to the current environmental conditions at every moment, a strategy based on quantum optimization is adopted to find the optimal weight value by measuring the quantum state. The quantum optimized Hamiltonian The calculation formula is: ， In the formula, is the feature dimension; is the importance parameter of the feature for regression prediction based on quantum Monte Carlo LSTM, and is the training parameter; Indicates that a specific feature Below, the output of the regression prediction function; S305. During the training process, the learning rate of LSTM is adaptively adjusted according to the feedback of the loss function of LSTM. LSTM based on quantum Monte Carlo can improve the convergence speed in the later stage of training and avoid premature convergence in the early stage. The adjustment method is expressed as: ， In the formula, is the learning rate of the LSTM at the k+1th iteration, is the learning rate of LSTM at the kth iteration, is the learning rate adjustment factor of LSTM, is the loss value of LSTM at the kth iteration, For LSTM The loss value of the iteration; S306. Repeat S302 to S305 until the preset stop iteration condition is met, which means that the LSTM training based on quantum Monte Carlo is completed.

7. The multi-mode integrated precipitation prediction method based on artificial intelligence according to claim 6 is characterized in that: Said S303 The calculation formula is: , In the formula, is the number of quantum sampling; is the quantum bit result in the quantum sampling process at the tth time step; is the activation function of LSTM, is the input feature of the i-th sample.

Images