System state evolution prediction method and device of complex system, and electronic equipment

Abstract

The invention provides a system state evolution prediction method and device of a complex system and electronic equipment, and relates to the technical field of state prediction, and the method comprises the steps: obtaining a historical system state sequence and a prediction moment of the complex system; performing dimension reduction processing on each element in the historical system state sequence to obtain a corresponding low-dimensional vector sequence; determining a matching code element vector matched with each low-dimensional vector in the low-dimensional vector sequence based on a plurality of set code element vectors to obtain a matching code element vector sequence; constructing a Voronoi diagram based on the matched code element vector sequence, and constructing a code element topological structure based on the Voronoi diagram; calculating an energy landscape based on the plurality of historical system states and the code element topology structure; determining a first matching code element vector at the prediction moment in the code element topological structure according to the energy landscape; and performing high-dimensional extension on the first matching code element vector to obtain a prediction system state corresponding to the prediction moment. According to the method, the accuracy of state prediction of the complex system can be effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of state prediction, and in particular, to a method and device for predicting the evolution of the system state of a complex system, and an electronic device. Background Art

[0002] Currently, complex systems are involved in many fields such as nature, biomedicine, industry, and society. For example, complex systems in the natural field include climate systems, ecological environment systems, ecological evolution systems, etc.; complex systems in the biomedicine field include biomolecular systems (such as protein folding), etc.; complex systems in the industrial field include power systems, etc.; complex systems in the social field include financial market systems, etc. The prediction of the system state of complex systems has potential important value in multiple practical application scenarios. For example, in the industrial field, predicting the change trend of the power grid operation in the power system can provide a scientific basis for power grid construction and power transmission management; in the field of intelligent manufacturing, it can help monitor the equipment state, predict possible failures, and improve the reliability and efficiency of industrial production; in the biomedicine field, it can be used to analyze the kinetic processes of biomolecules and provide theoretical support for drug development and disease research. Therefore, establishing a prediction model for the system state prediction of complex systems can be widely applied as a basic tool in many fields.

[0003] In actual situations, the number of variables in complex systems is numerous, which makes it difficult to analyze the complex systems in a high-dimensional state. At the same time, considering that the evolution of the state of complex systems is affected by multiple non-linear factors and randomness, currently, when facing the prediction of complex systems, related technologies all predict the system state of complex systems from aspects such as solving noise interference, increasing the understanding of dynamic multi-level systems, and analyzing the influence of features in complex systems on complex systems. Although the methods in related technologies can achieve the purpose of predicting the system state of complex systems to a certain extent, there are still great limitations in dealing with high-dimensional states, non-equilibrium behaviors, and the lack of sufficient labeled data, resulting in a low accuracy of the prediction results when predicting the evolution results of the system state of complex systems. Summary of the Invention

[0004] The present invention provides a method and device for predicting the evolution of the system state of a complex system, and an electronic device, so as to solve the defect of low accuracy of the prediction results in the prior art when predicting the evolution results of the system state of complex systems, and achieve the purpose of effectively improving the accuracy of the prediction results when predicting the evolution results of the system state of complex systems.

[0005] The present invention provides a method for predicting the evolution of the system state of a complex system, including the following steps.

[0006] Obtain the historical system state sequence of the complex system and obtain the prediction time; perform dimensionality reduction processing on each historical system state in the historical system state sequence to obtain the corresponding low-dimensional vector sequence of the historical system state sequence; for each low-dimensional vector in the low-dimensional vector sequence, determine the matching code element vector that matches each low-dimensional vector among the set of multiple code element vectors, and determine the corresponding matching code element vector sequence of the low-dimensional vector sequence based on the matching code element vectors that match each low-dimensional vector; construct a Voronoi diagram including all the matching code element vectors in the matching code element vector sequence based on the matching code element vector sequence, and construct a code element topological structure based on the Voronoi diagram; based on multiple historical system states and the code element topological structure, obtain the distribution of each matching code element vector in the matching code element vector sequence, and calculate the energy landscape of the complex system based on the distribution; according to the energy landscape, determine the first matching code element vector corresponding to the prediction time in the code element topological structure; perform high-dimensional expansion on the first matching code element vector to obtain the predicted system state corresponding to the prediction time.

[0007] According to a method for predicting the evolution of the system state of a complex system provided by the present invention, for each low-dimensional vector in the low-dimensional vector sequence, determining the matching code element vector that matches each low-dimensional vector among the set of multiple code element vectors includes: for each low-dimensional vector in the low-dimensional vector sequence, calculate its similarity with each code element vector among the multiple code element vectors to obtain multiple similarities corresponding one by one to the multiple code element vectors; determine the code element vector corresponding to the target similarity as the matching code element vector that matches; wherein, the target similarity is greater than the other similarities among the multiple similarities.

[0008] According to a method for predicting the evolution of the system state of a complex system provided by the present invention, according to the energy landscape, determining the first matching code element vector corresponding to the prediction time in the code element topological structure includes: according to the energy landscape, determine the target matching code element vector corresponding to the first system state at the current time in the code element topological structure, and determine the first transfer path from the current time corresponding to the first system state to the prediction time in the code element topological structure; starting from the position of the target matching code element vector in the code element topological structure, calculate the first state transfer probability of the target matching code element vector transferring to each matching code element vector in the code element topological structure along the first transfer path to obtain multiple first state transfer probabilities corresponding one by one to the multiple matching code element vectors in the code element topological structure; determine the matching code element vector corresponding to the target first state transfer probability as the first matching code element vector; wherein, the target first state transfer probability is greater than the other state transfer probabilities among the multiple state transfer probabilities.

[0009] According to a method for predicting the evolution of the system state of a complex system provided by the present invention, after determining the matching symbol vector corresponding to the target first state transition probability as the first matching symbol vector, the method further includes: obtaining parameter values of at least one state control parameter; predicting, according to the energy landscape and the parameter values of at least one state control parameter, a second transition path evolving from the current moment to the prediction moment in the symbol topology structure; taking the position of the target matching symbol vector in the symbol topology structure as the starting point, calculating the second state transition probability of the target matching symbol vector transferring to each matching symbol vector in the symbol topology structure along the second transition path, and obtaining a plurality of second state transition probabilities corresponding one by one to a plurality of matching symbol vectors in the symbol topology structure; determining the matching symbol vector corresponding to the target second state transition probability as the second matching symbol vector; wherein the target second state transition probability is greater than other second state transition probabilities among the plurality of second state transition probabilities; performing high-dimensional expansion on the second matching symbol vector to obtain a second predicted system state corresponding to the prediction moment.

[0010] According to a method for predicting the evolution of the system state of a complex system provided by the present invention, before obtaining the historical system state sequence of the complex system and before obtaining the prediction moment, the method further includes: establishing a target state transition prediction model based on a graph neural network and the Fokker-Planck equation; determining, according to the energy landscape, a target matching symbol vector corresponding to the first system state at the current moment in the symbol topology structure, and determining a first transition path evolving from the current moment corresponding to the first system state to the prediction moment in the symbol topology structure, including: inputting the energy landscape and the symbol topology structure into the target state transition prediction model, and obtaining the target matching symbol vector and the first transition path based on the target state transition prediction model.

[0011] According to a method for predicting the evolution of the system state of a complex system provided by the present invention, before obtaining the historical system state sequence of the complex system and before obtaining the prediction moment, the method further includes: establishing a target encoding model and a target decoding model coupled with the target encoding model; for each low-dimensional vector in the low-dimensional vector sequence, determining a matching symbol vector matching each low-dimensional vector among a set of multiple symbol vectors, and determining a matching symbol vector sequence corresponding to the low-dimensional vector sequence based on the matching symbol vectors matching each low-dimensional vector, including: inputting the multiple symbol vectors and the low-dimensional vector sequence into the target encoding model, and obtaining the matching symbol vector sequence based on the target encoding model; performing high-dimensional expansion on the first matching symbol vector to obtain a predicted system state corresponding to the prediction moment, including: inputting the first matching symbol vector into the target decoding model, and obtaining the predicted system state based on the target decoding model.

[0012] A method for predicting the evolution of the system state of a complex system provided by the present invention includes establishing a target encoding model and a target decoding model coupled with the target encoding model, which includes jointly training an initial encoding model, an initial decoding model, and an initial state transition prediction model to obtain a target encoding model, a target decoding model, and a target state transition prediction model; wherein, the first loss function of the initial encoding model and the initial decoding model is the two-norm distance between the first input vector input to the initial encoding model and the second output vector output by the initial decoding model; the second loss function of the initial state transition prediction model is the two-norm distance between the second input vector and the second output vector of the initial state transition model.

[0013] The present invention also provides a device for predicting the evolution of the system state of a complex system, including the following modules: an acquisition module, a dimensionality reduction module, a determination module, a processing module, a calculation module, a prediction module, and a high-dimensional expansion module.

[0014] The acquisition module is used to acquire the historical system state sequence of the complex system and the prediction time.

[0015] The dimensionality reduction module is used to perform dimensionality reduction processing on each historical system state in the historical system state sequence to obtain a low-dimensional vector sequence corresponding to the historical system state sequence.

[0016] The determination module is used to determine, for each low-dimensional vector in the low-dimensional vector sequence, the matching code element vector that matches each low-dimensional vector among a set of multiple code element vectors, and determine the matching code element vector sequence corresponding to the low-dimensional vector sequence based on the matching code element vectors that match each low-dimensional vector.

[0017] The processing module is used to construct a Voronoi diagram including all the matching code element vectors in the matching code element vector sequence based on the matching code element vector sequence, and construct a code element topological structure based on the Voronoi diagram.

[0018] The calculation module is used to obtain the distribution of each matching code element vector in the matching code element vector sequence based on a plurality of historical system states and the code element topological structure, and calculate the energy landscape of the complex system based on the distribution.

[0019] The prediction module is used to determine the first matching code element vector corresponding to the prediction time in the code element topological structure according to the energy landscape.

[0020] The high-dimensional expansion module is used to perform high-dimensional expansion on the first matching code element vector to obtain the predicted system state corresponding to the prediction time.

[0021] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, it implements the method for predicting the evolution of the system state of a complex system as described above.

[0022] The present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the state evolution prediction method of any one of the above complex systems.

[0023] The present invention also provides a computer program product, including a computer program, and when the computer program is executed by a processor, it implements the state evolution prediction method of any one of the above complex systems.

[0024] The method and device for predicting the state evolution of a complex system provided by the present invention, as well as an electronic device, obtain the historical system state sequence of the complex system and the prediction time; perform dimensionality reduction processing on each historical system state in the historical system state sequence to obtain a low-dimensional vector sequence corresponding to the historical system state sequence; for each low-dimensional vector in the low-dimensional vector sequence, determine the matching code element vector that matches each low-dimensional vector among a set of multiple code element vectors, and determine the matching code element vector sequence corresponding to the low-dimensional vector sequence based on the matching code element vectors that match each low-dimensional vector; construct a Voronoi diagram including all the matching code element vectors in the matching code element vector sequence based on the matching code element vector sequence, and construct a code element topological structure based on the Voronoi diagram; calculate the energy landscape of the complex system based on the code element topological structure and the distribution of each matching code element vector in the matching code element vector sequence; determine the first matching code element vector corresponding to the prediction time in the code element topological structure according to the energy landscape; perform high-dimensional expansion on the first matching code element vector to obtain the predicted system state corresponding to the prediction time. It can be seen that the present invention can perform dimensionality reduction processing on the historical system state sequence of the complex system with high-dimensional states to obtain the corresponding low-dimensional vector sequence, discretize the low-dimensional vector sequence based on a set of multiple code element vectors to obtain a discrete matching code element vector sequence, construct a Voronoi diagram based on the matching code element vector sequence, construct a code element topological structure based on the Voronoi diagram, calculate the energy landscape of the complex system based on the code element topological structure and the distribution of each matching code element vector in the matching code element vector sequence. In the present invention, the acquisition of the energy landscape does not require additional acquisition of sample annotation data for the energy landscape, which can not only efficiently estimate the energy landscape of the complex system, but also effectively overcome the limitations brought by the lack of sufficient annotation data to the prediction results; then, according to the energy landscape, determine the first matching code element vector corresponding to the prediction time in the code element topological structure to effectively address the limitations brought by non-equilibrium behavior to the prediction of the complex system. Finally, perform high-dimensional expansion on the low-dimensional space vector corresponding to the first matching code element vector to obtain the predicted system state at the prediction time, realizing accurate prediction of the system state evolution result of the complex system and effectively improving the accuracy of predicting the system state evolution result of the complex system. At the same time, in the process of realizing the prediction of the system state, the present invention only needs to obtain multiple historical system states corresponding to multiple historical times, which can effectively reduce the data acquisition cost. Description of the Drawings

[0025] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0026] Figure 1 It is one of the schematic flowcharts of the method for predicting the state evolution of a complex system provided by the present invention.

[0027] Figure 2 It is a schematic diagram of the historical observation trajectory corresponding to the historical system state sequence provided by the present invention, and the actual energy landscape of the historical observation trajectory.

[0028] Figure 3 It is a schematic diagram of obtaining a symbol topological structure based on a low-dimensional vector sequence provided by the present invention.

[0029] Figure 4 It is a schematic diagram of obtaining the predicted system state at the prediction moment based on the symbol topological structure provided by the present invention.

[0030] Figure 5 It is a schematic structural diagram of the device for predicting the state evolution of a complex system provided by the present invention.

[0031] Figure 6 It is a schematic structural diagram of the electronic device provided by the present invention. Detailed implementation manners

[0032] To make the objectives, technical solutions and advantages of the present invention clearer, the technical solutions in the present invention will be clearly and completely described below with reference to the accompanying drawings in the present invention. Obviously, the described embodiments are some but not all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art without creative efforts based on the embodiments in the present invention belong to the scope of protection of the present invention.

[0033] In the related art, the energy landscape can describe the stability of each system state of a complex system and the conversion paths between different system states, and provide a basis for predicting the system state of the complex system.

[0034] In related technologies, most of the system state prediction methods for complex systems are based on statistical models and machine learning models. Statistical models such as Markov state models and classical time series analysis methods attempt to predict the future evolution state of the system by assuming that the system satisfies specific equilibrium states and linear relationships. However, complex systems usually exhibit non-equilibrium states and highly non-linear characteristics, which makes traditional statistical models show obvious deficiencies when dealing with complex environments. In addition, traditional statistical methods also rely on the collection of a large amount of historical data, with high data acquisition costs and limited coverage, and it is difficult to capture mutations or rare events in the system. In recent years, deep learning technologies have been applied to the state prediction of complex systems. These methods can automatically extract features from a large amount of historical data and use them to train prediction models, thereby improving prediction accuracy. However, the prediction models based on deep learning rely on a large amount of labeled data, and the models often lack physical meaning interpretability, making it difficult to be verified and trusted in practical applications. At the same time, although self-supervised learning reduces the need for labeled data, it lacks an understanding of physical mechanisms and is difficult to effectively capture the energy landscape information of the system, affecting the accuracy of prediction results.

[0035] Based on the above existing problems, an embodiment of the present invention provides a method for predicting the system state evolution of a complex system to effectively solve the problem of low prediction result accuracy in predicting the system state evolution result of a complex system in the prior art.

[0036] The following combines Figures 1 - 4 to describe the method for predicting the system state evolution of a complex system provided by an embodiment of the present invention.

[0037] Figure 1 is one of the flow diagrams of the method for predicting the system state evolution of a complex system provided by an embodiment of the present invention. As Figure 1 shown, the method includes the following S110~S170.

[0038] S110: Obtain the historical system state sequence of the complex system and obtain the prediction moment.

[0039] The complex system in the embodiment of the present invention refers to a non-linear system composed of multiple interacting components, and its evolution has high complexity and randomness.

[0040] For the complex system in the embodiment of the present invention, let the system state of the complex system be , and the dynamic equation of the system state of the complex system can be expressed as: . Among them, represents the deterministic evolution part of the complex system, represents the random perturbation, It is a Wiener process. In specific implementations, complex systems that satisfy the above kinetic equations widely exist in the natural, industrial, and social fields, and usually exhibit characteristics of non-equilibrium, multi-scale, and high non-linearity, making it difficult to predict through simple linear models.

[0041] The historical system state sequence includes multiple historical system states corresponding to multiple historical moments one by one, and the multiple historical system states in the historical system state sequence are arranged in the order of the historical moments.

[0042] The prediction moment is a moment that has not yet evolved in the continuous evolution of the complex system. For example, when a user wants to know the system state of the complex system at a certain future time point, this time point is the prediction moment. The prediction moment can be set by the user according to needs.

[0043] As Figure 2 shown, based on the historical system state sequence, a historical observation trajectory can be depicted. Based on the historical observation trajectory, the energy at each position point of the historical observation trajectory is different. Describing these energies is the energy landscape.

[0044] Specifically, the energy landscape is used to describe the stability and transition paths of each system state in the complex system. Let the energy function of the complex system be ， Then the energy landscape is defined as the energy level corresponding to each system state in the system state space. As Figure 2 shown, the low-energy region represents the stable state of the complex system, and the high-energy region corresponds to the unstable or highly uncertain system state. The state probability distribution of the complex system in the thermal equilibrium state can be expressed as: . Among them, is the Boltzmann constant, is the temperature of the complex system. The construction of the energy landscape helps to understand the dynamic behavior of the system, reveals the mechanism of the complex system's transition between different states, and thus provides a theoretical basis for system state prediction.

[0045] Please continue to refer to Figure 2 , based on the historical system state sequence, a historical observation trajectory can be obtained. The energy at each trajectory position is different. Establish a three-dimensional coordinate ( Figure 2 not shown in the figure) based on the plane where the historical observation trajectory is located. The coordinate axis perpendicular to this plane is used to represent the energy size at each position on the historical observation trajectory, and the energy landscape as shown in Figure 2 can be obtained. Figure 2 The energy landscape shown in the figure needs to be obtained by collecting energy data from reality, and the data acquisition is difficult. To solve this problem, in the embodiments of the present invention, the energy landscape of the complex system can be estimated based on the historical system state sequence, effectively reducing the difficulty of data acquisition and being able to obtain an accurate energy landscape.

[0046] Next, the process of constructing the energy landscape in the embodiments of the present invention will be introduced.

[0047] S120: Perform dimensionality reduction processing on each historical system state in the historical system state sequence to obtain a low-dimensional vector sequence corresponding to the historical system state sequence.

[0048] Each historical system state in the historical system state sequence can be encoded by a neural network encoder to obtain a low-dimensional vector corresponding to each historical system state, and based on the low-dimensional vectors corresponding to each historical system state, a low-dimensional vector sequence is obtained. The order of any low-dimensional vector in the low-dimensional vector sequence is the same as the order of the historical system state corresponding to the low-dimensional vector in the historical system state sequence.

[0049] Exemplarily, if the historical system state sequence is , where T is the total number of elements in the historical system state sequence, and T is a positive integer. For encoding, convert to a vector, and here use the intermediate latent vector to represent the vector obtained after encoding , and this vector is the state vector corresponding to the historical system state . Among them, , , and m is the dimension of each low-dimensional vector in the low-dimensional vector sequence.

[0050] S130: For each low-dimensional vector in the low-dimensional vector sequence, determine the matching code element vector that matches each low-dimensional vector among the set of multiple code element vectors, and based on the matching code element vectors that match each low-dimensional vector, determine the matching code element vector sequence corresponding to the low-dimensional vector sequence.

[0051] The multiple code element vectors can be all the code element vectors included in the set code table. The set code table can be a custom code table or a code table in an existing standard or specification.

[0052] Exemplarily, the code table where K is the number of code element vectors in the code table, and K is a positive integer. , , ……, are respectively the code element vectors in the code table .

[0053] When determining the matching symbol vectors that match each low-dimensional vector among a set of preset symbol vectors, for each low-dimensional vector in the low-dimensional vector sequence, the similarity between it and each symbol vector among the multiple symbol vectors can be calculated to obtain multiple similarities corresponding one by one to the multiple symbol vectors; then the symbol vector corresponding to the target similarity is determined as the matching symbol vector that matches the low-dimensional vector; where the target similarity is greater than the other similarities among the multiple similarities, that is, the maximum similarity among the multiple similarities is determined as the target similarity. Based on the matching symbol vectors that match each low-dimensional vector, a matching symbol vector sequence is obtained. Among them, the order of each matching symbol vector in the matching symbol vector sequence is the same as the order of the low-dimensional vector corresponding to the matching symbol vector in the low-dimensional vector sequence.

[0054] In some embodiments, the above similarity can be measured by the two-norm distance. The smaller the two-norm distance, the greater the similarity.

[0055] Taking the above example as an example, the code table , for the state vector obtained after encoding is . In this case, for any target symbol vector in the code table , calculate and 's two-norm distance , and the following multiple two-norm distances are obtained: , , ……, . Among the above multiple two-norm distances, select the smallest two-norm distance , from which it can be determined that has the greatest similarity with . Take as the target symbol vector, , where k is a positive integer. Finally, assign the value of to to obtain the matching symbol vector sequence corresponding to the low-dimensional vector sequence.

[0056] Of course, it can be understood that in the above example, only an exemplary way of measuring the similarity between the target state vector and each symbol vector among the multiple symbol vectors by the two-norm distance is given. The ways of measuring the similarity between the target state vector and each symbol vector among the multiple symbol vectors in the embodiments of the present invention include but are not limited to the two-norm distance. For example, it can also be cosine similarity, Euclidean distance, Manhattan distance, and so on.

[0057] In specific implementation, the execution process of S130 can be realized by an autoencoder. The autoencoder can be, for example, a Vector Quantized Variational Autoencoder (VQ-VAE). As Figure 2 shown, based on the target encoding model, the low-dimensional vector sequence (which can also be regarded as the low-dimensional vector space of the energy landscape) is encoded to obtain a discretized matching code element vector sequence.

[0058] The autoencoder in the embodiment of the present invention can be obtained based on model training. For the specific training method, refer to the introduction of the training part of the target encoding model and the target decoding model in the following text, and details are not described here.

[0059] It should be noted that through S120~S130, not only can the representation dimension of the system state of a complex system be effectively reduced, but also the main dynamic patterns of the complex system can be captured through discretized encoding (the low-dimensional vector sequence is a continuous sequence, and the matching code element vector sequence obtained based on the low-dimensional vector sequence and a set of code element vectors is the discretized encoding of the low-dimensional vector sequence). Thus, while maintaining the dynamic characteristics of the complex system, the computational complexity is greatly reduced, and the efficiency of modeling and prediction is improved. Through discretized encoding, the low-dimensional vector sequence can be quantized to effectively reduce the dimension of the system state of the complex system, and at the same time, it provides a reliable basis for the calculation of the energy landscape in the subsequent step (corresponding to S150) and the prediction of the system state in the subsequent steps (corresponding to S160~S170).

[0060] S140: Based on the matching code element vector sequence, construct a Voronoi diagram that includes all the matching code element vectors in the matching code element vector sequence, and construct a code element topological structure based on the Voronoi diagram.

[0061] The matching code element vector sequence is a set of discrete vectors. Therefore, a Voronoi diagram can be directly generated based on all the matching code element vectors in the matching code element vector sequence. The Voronoi diagram includes multiple Voronoi cells, and each Voronoi cell corresponds to a matching code element vector in the matching code element vector sequence. Each matching code element vector in the matching code element vector sequence is the seed point of the Voronoi diagram.

[0062] Exemplarily, the Voronoi diagram divides the vector space corresponding to the low-dimensional vector sequence into several regions, and the vectors in each region are assigned to the matching code element vector closest to them, thereby defining the division of the system state of the complex system.

[0063] Based on the positional relationship of the matching code element vectors in the Voronoi diagram, construct a code element topological structure The symbol topology structure can be referred to the corresponding part shown in Figure 3 .

[0064] S150: Based on multiple historical system states and the symbol topology structure, obtain the distribution of each matching symbol vector in the matching symbol vector sequence, and calculate the energy landscape of the complex system based on the distribution.

[0065] In the embodiment of the present invention, based on the matching symbol vectors included in the symbol topology structure, count the matching symbol vectors corresponding to each historical system state in multiple historical system states, obtain the data set distribution p of each matching symbol vector in multiple historical system states, input p into the energy prediction model, and calculate the energy of each matching symbol vector based on the energy prediction model , based on calculate the Boltzmann distribution as the distribution q of each matching symbol vector in the matching symbol vector sequence, and use the KL (Kullback-Leibler) divergence between p and q as the loss function of the physically guided regularization term , based on it can be required that the energy predicted by the energy prediction model matches the true distribution of the data.

[0066] Calculate the energy of each matching symbol vector based on the above process , based on the energy of each matching symbol vector , construct the energy landscape of the complex system .

[0067] S160: Determine the first matching symbol vector corresponding to the prediction moment in the symbol topology structure according to the energy landscape.

[0068] The target state transition prediction model can be established in advance based on the Graph Neural Network (GNN) and the Graph Neural Fokker-Planck Equation. The target state transition prediction model includes the following formulas (1) - (3).

[0069] Formula (1): .

[0070] Formula (2): .

[0071] Formula (3): .

[0072] In formula (1), encodes the initial condition, is a D-dimensional one-hot vector, representing the probability density at the current moment, and the th element is 1.

[0073] In formula (3), , is the weight obtained by neighborhood attention calculation; is the learning coefficient, representing the intensity of the noise effect between adjacent symbols; is the scaling factor of the sigmoid activation function.

[0074] When establishing a target state transition prediction model based on a graph neural network and the Fokker - Planck equation, a graph neural network can be used to model the state transition process of a complex system in a low - dimensional vector. The loss function in this process is . The graph neural network can learn the dependency relationship between symbols through a multi - layer propagation mechanism, thereby effectively capturing the dynamic characteristics of the system state. During the training process, a self - supervised learning method is adopted, enabling the model to still have good prediction performance in the absence of a large amount of labeled data.

[0075] Based on the established target state transition prediction model above, the energy landscape and the symbol topology structure can be input into the target state transition prediction model. Based on the target state transition prediction model, the target matching symbol vector corresponding to the first system state at the current moment can be determined in the symbol topology structure. As Figure 3 shown, based on the probability density at the current moment, the target matching symbol vector corresponding to the first system state at the current moment is determined as D1. After passing through the Fokker - Planck equation (i.e., the above formulas (2) - (3)), the first transition path from the current moment corresponding to the first system state to the prediction moment is determined in the symbol topology structure (such as Figure 4 L in). Then, starting from the position of the target matching symbol vector in the symbol topology structure, the first state transition probability of the target matching symbol vector along the first transition path to each matching symbol vector in the symbol topology structure is calculated, obtaining multiple first state transition probabilities corresponding to multiple matching symbol vectors in the symbol topology structure; finally, the matching symbol vector corresponding to the target first state transition probability is determined as the first matching symbol vector; where the target first state transition probability is greater than other state transition probabilities among the multiple state transition probabilities. That is to say, the largest first state transition probability among the multiple first state transition probabilities is selected as the target first state transition probability. The first matching symbol vector is like Figure 4 D2 in.

[0076] In some embodiments, in order to observe the influence of some state control parameters on the state of a complex system, after determining the matching symbol vector corresponding to the target first state transition probability as the first matching symbol vector, it is also possible to obtain the parameter values of at least one state control parameter. For example, for protein folding, the temperature, environmental pH value, etc. set by the user can be obtained. The at least one state control parameter and its parameter values can be set by those skilled in the art according to the actual situation.

[0077] After obtaining the parameter values of at least one state control parameter, according to the energy landscape and the parameter values of at least one state control parameter, predict the second transition path from the current moment to the prediction moment in the symbol topology structure. The influence factor of at least one state control parameter is added to the prediction of the second transition path. In this case, starting from the position of the target matching symbol vector in the symbol topology structure, calculate the second state transition probability of the target matching symbol vector transferring to each matching symbol vector in the symbol topology structure along the second transition path, and obtain multiple second state transition probabilities corresponding one by one to the multiple matching symbol vectors in the symbol topology structure. Determine the matching symbol vector corresponding to the target second state transition probability as the second matching symbol vector; where the target second state transition probability is greater than the other second state transition probabilities among the multiple second state transition probabilities.

[0078] In some embodiments, a comparison graph of the second matching symbol vector and the first matching symbol vector can be output for the user to analyze relevant situations such as the influence result and influence degree of at least one state control parameter on the complex system, and provide auxiliary information for the user to analyze the evolution process of the complex system. For example, it can be known which state control parameters can make the system state of the complex system evolve in the desired direction, and which state control parameters can make the system state of the complex system evolve in an undesired direction, so as to suppress the undesired parameter combinations and promote the complex system to develop along the predetermined evolution direction.

[0079] In the embodiments of the present invention, based on the graph neural network and S170: Perform high-dimensional expansion on the first matching symbol vector to obtain the predicted system state corresponding to the prediction moment.

[0080] As Figure 4 shown, the target decoding model coupled with the target encoding model can be used to perform high-dimensional expansion on the first matching symbol vector D2, that is, perform system state reconstruction on the first matching symbol vector D2 in the low-dimensional state to obtain the system state in the original high-dimensional state of the complex system corresponding to the first matching symbol vector. This system state is the predicted system state, and finally the prediction of the system state at the prediction moment is realized.

[0081] The following is a description of the process of acquiring the target encoding model, the target decoding model, and the target state transition prediction model in the embodiment of the present invention.

[0082] In an embodiment of the present invention, before executing S110, an initial coding model, an initial decoding model and an initial state transfer model can be obtained in advance, and then the initial coding model, the initial decoding model and the initial state transfer model are jointly trained to obtain a target coding model, a target decoding model and a target state transfer prediction model; wherein the first loss function of the initial coding model and the initial decoding model is the binomial distance between the first input vector of the initial coding model and the second output vector of the initial decoding model; and the second loss function of the initial state transfer model is the binomial distance between the second input vector and the second output vector of the initial state transfer model.

[0083] Through joint training, the target encoding model, target decoding model and target state transfer model are optimized collaboratively to achieve high-precision prediction of the state of complex systems.

[0084] Two specific embodiments are listed below.

[0085] Embodiment 1 The complex system is protein folding. In order to obtain the prediction of the folding state of protein folding, the historical folding state data of protein folding collected in advance by the user can be first obtained, and then the historical folding state data can be reduced in dimension (for example, the historical folding state data can be reduced in dimension based on a neural network encoder), and the molecular conformation in a high-dimensional state can be mapped to a low-dimensional state to obtain a low-dimensional vector sequence. The low-dimensional vector sequence is discretized using VQ-VAE to obtain a matching code element vector sequence, and each code element in the matching code element vector sequence represents a different folding state.

[0086] Next, we construct an energy landscape and Voronoi diagram based on the matching code element vector sequence, and use the Fokker-Planck equation to describe the state transition probability between different conformations during protein folding. We train the graph neural network (GNN) to learn the transition relationship between different folding states, thereby modeling the dynamic behavior of the protein folding process.

[0087] Finally, by combining the deep learning model with the physical-guided regularization term, the present invention can achieve high-precision prediction of the future folding state of proteins and identify the key states and unstable intermediate states that may occur in the protein folding process. It effectively solves the nonlinear and random problems in the protein folding process, successfully predicts the future structural changes of protein molecules, and provides an important auxiliary tool for the structural prediction of biomolecules and drug design.

[0088] Embodiment 2 The complex system is an ecological evolutionary system. An ecological evolutionary system usually involves the interactions among multiple species and complex changes in the environment. Its dynamic evolution process is affected by various internal and external factors and is difficult to be effectively controlled by traditional methods. In the embodiments of the present invention, the dynamic data of species populations in the ecological evolutionary system pre-collected by a user can be obtained, and then the dynamic data of species populations is dimensionally reduced, and the high-dimensional species interaction representation is mapped to a low dimension to obtain a low-dimensional vector sequence. The low-dimensional vector sequence is discretized by VQ-VAE to obtain a matching code element vector sequence, and each code element in the matching code element vector sequence represents a different ecological equilibrium state (or non-equilibrium state).

[0089] Next, an energy landscape and a Voronoi diagram are constructed based on the matching code element vector sequence, and the Fokker-Planck equation is used to describe the state transition process of the interactions among species in the ecological evolutionary system.

[0090] In the embodiments of the present invention, an expected position can also be found on the energy landscape, the path between the system state at the current moment and the system state at the prediction moment of the ecological evolutionary system can be located, and a control signal (i.e., the parameter value of at least one state control parameter mentioned above) can be designed to observe the evolution path of the ecological evolutionary system, and the parameter value of the state control parameter that can make the evolution path meet the expectation or the parameter value of the state control parameter that can make the evolution path not meet the expectation can be found.

[0091] For example, by adjusting the population quantity of a specific species, the suppression of an undesired species combination or the avoidance of ecological imbalance can be achieved, so as to promote the ecological system to develop along a predetermined evolution direction.

[0092] Based on the method in the embodiments of the present invention, the accurate prediction and directional regulation of the future state of the ecological evolutionary system can be realized to maintain the health and stability of the ecological evolutionary system. This has significant value for ecological protection, species diversity management and their applications. By designing a control signal to make the ecological evolutionary system evolve along the desired ecological evolution path or timely discovering the factors that cause the ecological evolutionary system to evolve along an undesired ecological evolution path, the occurrence of an undesired species combination or ecological imbalance can be avoided. Realizing the precise control and directional evolution of the future state of the ecological system, and further promoting the healthy and stable development of the ecological system, this has important application value for ecological protection and species diversity management.

[0093] The method for predicting the state evolution of a complex system provided by the present invention includes: obtaining the historical system state sequence of the complex system and the prediction time; performing dimensionality reduction processing on each historical system state in the historical system state sequence to obtain a low-dimensional vector sequence corresponding to the historical system state sequence; for each low-dimensional vector in the low-dimensional vector sequence, determining the matching code element vector that matches each low-dimensional vector among a set of multiple code element vectors, and determining a matching code element vector sequence corresponding to the low-dimensional vector sequence based on the matching code element vectors that match each low-dimensional vector; constructing a Voronoi diagram including all the matching code element vectors in the matching code element vector sequence based on the matching code element vector sequence, and constructing a code element topological structure based on the Voronoi diagram; calculating the energy landscape of the complex system based on the code element topological structure and the distribution of each matching code element vector in the matching code element vector sequence; determining the first matching code element vector corresponding to the prediction time in the code element topological structure according to the energy landscape; and performing high-dimensional expansion on the first matching code element vector to obtain the predicted system state corresponding to the prediction time. It can be seen that the present invention can perform dimensionality reduction processing on the historical system state sequence of the complex system with high-dimensional states to obtain the corresponding low-dimensional vector sequence, discretize the low-dimensional vector sequence based on a set of multiple code element vectors to obtain a discrete matching code element vector sequence, construct a Voronoi diagram based on the matching code element vector sequence, construct a code element topological structure based on the Voronoi diagram, calculate the energy landscape of the complex system based on the code element topological structure and the distribution of each matching code element vector in the matching code element vector sequence. In the present invention, the acquisition of the energy landscape does not require additional acquisition of sample annotation data for the energy landscape, which can not only efficiently estimate the energy landscape of the complex system, but also effectively overcome the limitations brought to the prediction result due to the lack of sufficient annotation data; then, according to the energy landscape, determine the first matching code element vector corresponding to the prediction time in the code element topological structure to effectively cope with the limitations brought by non-equilibrium behavior to the prediction of the complex system. Finally, perform high-dimensional expansion on the low-dimensional space vector corresponding to the first matching code element vector to obtain the predicted system state at the prediction time, realizing accurate prediction of the system state evolution result of the complex system and effectively improving the accuracy of predicting the system state evolution result of the complex system. At the same time, in the process of realizing the prediction of the system state, the present invention only needs to obtain multiple historical system states corresponding to multiple historical times, which can effectively reduce the data acquisition cost.

[0094] The state evolution prediction device of the complex system provided by the present invention will be described below. The state evolution prediction device of the complex system described below can be correspondingly referred to the state evolution prediction method of the complex system described above.

[0095] Figure 5 is a schematic structural diagram of the state evolution prediction device of the complex system provided by the present invention. As Figure 5As shown in the figure, the prediction of the state evolution of a complex system 500 includes: an acquisition module 501, a dimensionality reduction module 502, a determination module 503, a processing module 504, a calculation module 505, a prediction module 506, and a high-dimensional expansion module 507.

[0096] The acquisition module 501 is configured to acquire the historical system state sequence of the complex system and acquire the prediction time.

[0097] The dimensionality reduction module 502 is configured to perform dimensionality reduction processing on each historical system state in the historical system state sequence to obtain a low-dimensional vector sequence corresponding to the historical system state sequence.

[0098] The determination module 503 is configured to determine, for each low-dimensional vector in the low-dimensional vector sequence, a matching code element vector that matches each low-dimensional vector among a set of multiple code element vectors, and determine a matching code element vector sequence corresponding to the low-dimensional vector sequence based on the matching code element vectors that match each low-dimensional vector.

[0099] The processing module 504 is configured to construct a Voronoi diagram including all the matching code element vectors in the matching code element vector sequence based on the matching code element vector sequence, and construct a code element topological structure based on the Voronoi diagram.

[0100] The calculation module 505 is configured to obtain the distribution of each matching code element vector in the matching code element vector sequence based on a plurality of historical system states and the code element topological structure, and calculate the energy landscape of the complex system based on the distribution.

[0101] The prediction module 506 is configured to determine a first matching code element vector corresponding to the prediction time in the code element topological structure according to the energy landscape.

[0102] The high-dimensional expansion module 507 is configured to perform high-dimensional expansion on the first matching code element vector to obtain a predicted system state corresponding to the prediction time.

[0103] The state evolution prediction device for a complex system provided by the present invention obtains the historical system state sequence of the complex system and the prediction time; performs dimensionality reduction processing on each historical system state in the historical system state sequence to obtain a low-dimensional vector sequence corresponding to the historical system state sequence; for each low-dimensional vector in the low-dimensional vector sequence, determines the matching code element vector that matches each low-dimensional vector among a set of multiple code element vectors, and determines the matching code element vector sequence corresponding to the low-dimensional vector sequence based on the matching code element vectors that match each low-dimensional vector; constructs a Voronoi diagram including all the matching code element vectors in the matching code element vector sequence based on the matching code element vector sequence, and constructs a code element topological structure based on the Voronoi diagram; calculates the energy landscape of the complex system based on the code element topological structure and the distribution of each matching code element vector in the matching code element vector sequence; determines the first matching code element vector corresponding to the prediction time in the code element topological structure according to the energy landscape; performs high-dimensional expansion on the first matching code element vector to obtain the predicted system state corresponding to the prediction time. It can be seen that the present invention can perform dimensionality reduction processing on the historical system state sequence of the complex system with high-dimensional states to obtain the corresponding low-dimensional vector sequence, discretize the low-dimensional vector sequence based on a set of multiple code element vectors to obtain a discrete matching code element vector sequence, construct a Voronoi diagram based on the matching code element vector sequence, construct a code element topological structure based on the Voronoi diagram, calculate the energy landscape of the complex system based on the code element topological structure and the distribution of each matching code element vector in the matching code element vector sequence. In the present invention, the acquisition of the energy landscape does not require additional acquisition of sample annotation data for the energy landscape, which can not only efficiently estimate the energy landscape of the complex system, but also effectively overcome the limitations brought to the prediction results due to the lack of sufficient annotation data; then, according to the energy landscape, the first matching code element vector corresponding to the prediction time is determined in the code element topological structure to effectively address the limitations brought by non-equilibrium behavior to the prediction of the complex system. Finally, the low-dimensional space vector corresponding to the first matching code element vector is subjected to high-dimensional expansion to obtain the predicted system state at the prediction time, realizing accurate prediction of the system state evolution result of the complex system and effectively improving the accuracy of predicting the system state evolution result of the complex system. At the same time, in the process of realizing the prediction of the system state, the present invention only needs to obtain multiple historical system states corresponding to multiple historical times, which can effectively reduce the data acquisition cost.

[0104] Figure 6 Illustrates a schematic physical structure diagram of an electronic device, such as Figure 6As shown in the figure, the electronic device may include: a processor 610, a communications interface 620, a memory 630, and a communication bus 640. Among them, the processor 610, the communications interface 620, and the memory 630 complete their mutual communication through the communication bus 640. The processor 610 may call the logical instructions in the memory 630 to execute a method for predicting the state evolution of a complex system. The method includes: obtaining a historical system state sequence of the complex system and obtaining a prediction time; performing dimensionality reduction processing on each historical system state in the historical system state sequence to obtain a low-dimensional vector sequence corresponding to the historical system state sequence; for each low-dimensional vector in the low-dimensional vector sequence, determining a matching code element vector that matches each low-dimensional vector among a set of multiple code element vectors, and determining a matching code element vector sequence corresponding to the low-dimensional vector sequence based on the matching code element vectors that match each low-dimensional vector; constructing a Voronoi diagram including all the matching code element vectors in the matching code element vector sequence based on the matching code element vector sequence, and constructing a code element topological structure based on the Voronoi diagram; obtaining the distribution of each matching code element vector in the matching code element vector sequence based on multiple historical system states and the code element topological structure, and calculating the energy landscape of the complex system based on the distribution; determining a first matching code element vector corresponding to the prediction time in the code element topological structure according to the energy landscape; performing high-dimensional expansion on the first matching code element vector to obtain a predicted system state corresponding to the prediction time.

[0105] In addition, when the logical instructions in the above-mentioned memory 630 are implemented in the form of software functional units and sold or used as an independent product, they may be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, may be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The foregoing storage medium includes: various media such as a USB flash drive, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk, or an optical disc that can store program codes.

[0106] On the other hand, the present invention also provides a computer program product, which includes a computer program. The computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the state evolution prediction method of the complex system provided by each of the above methods. The method includes: obtaining a historical system state sequence of the complex system and obtaining a prediction time; performing dimensionality reduction processing on each historical system state in the historical system state sequence to obtain a low-dimensional vector sequence corresponding to the historical system state sequence; for each low-dimensional vector in the low-dimensional vector sequence, determining a matching code element vector that matches each low-dimensional vector among a set of multiple code element vectors, and determining a matching code element vector sequence corresponding to the low-dimensional vector sequence based on the matching code element vectors that match each low-dimensional vector; constructing a Voronoi diagram including all the matching code element vectors in the matching code element vector sequence based on the matching code element vector sequence, and constructing a code element topological structure based on the Voronoi diagram; obtaining the distribution of each matching code element vector in the matching code element vector sequence based on multiple historical system states and the code element topological structure, and calculating the energy landscape of the complex system based on the distribution; determining a first matching code element vector corresponding to the prediction time in the code element topological structure according to the energy landscape; and performing high-dimensional expansion on the first matching code element vector to obtain a predicted system state corresponding to the prediction time.

[0107] On the other hand, the present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it is implemented to execute the state evolution prediction method of the complex system provided by each of the above methods. The method includes: obtaining a historical system state sequence of the complex system and obtaining a prediction time; performing dimensionality reduction processing on each historical system state in the historical system state sequence to obtain a low-dimensional vector sequence corresponding to the historical system state sequence; for each low-dimensional vector in the low-dimensional vector sequence, determining a matching code element vector that matches each low-dimensional vector among a set of multiple code element vectors, and determining a matching code element vector sequence corresponding to the low-dimensional vector sequence based on the matching code element vectors that match each low-dimensional vector; constructing a Voronoi diagram including all the matching code element vectors in the matching code element vector sequence based on the matching code element vector sequence, and constructing a code element topological structure based on the Voronoi diagram; obtaining the distribution of each matching code element vector in the matching code element vector sequence based on multiple historical system states and the code element topological structure, and calculating the energy landscape of the complex system based on the distribution; determining a first matching code element vector corresponding to the prediction time in the code element topological structure according to the energy landscape; and performing high-dimensional expansion on the first matching code element vector to obtain a predicted system state corresponding to the prediction time.

[0108] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. A person of ordinary skill in the art can understand and implement it without creative work.

[0109] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution, in essence, or the part that contributes to the prior art can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0110] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features. And these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the evolution of the system state of a complex system, characterized in that Including: Obtaining a historical system state sequence of a complex system and obtaining a prediction time; Performing dimensionality reduction processing on each historical system state in the historical system state sequence to obtain a low-dimensional vector sequence corresponding to the historical system state sequence; For each low-dimensional vector in the low-dimensional vector sequence, determining a matching code element vector that matches each low-dimensional vector among a set of multiple code element vectors, and determining a matching code element vector sequence corresponding to the low-dimensional vector sequence based on the matching code element vectors that match each low-dimensional vector; Constructing a Voronoi diagram including all the matching code element vectors in the matching code element vector sequence based on the matching code element vector sequence, and constructing a code element topological structure based on the Voronoi diagram; Based on the multiple historical system states and the code element topological structure, obtaining the distribution of each matching code element vector in the matching code element vector sequence, and calculating the energy landscape of the complex system based on the distribution; According to the energy landscape, determining a first matching code element vector corresponding to the prediction time in the code element topological structure; Performing high-dimensional expansion on the first matching code element vector to obtain a predicted system state corresponding to the prediction time.

2. The method for predicting the system state evolution of a complex system according to claim 1, characterized in that The step of, for each low-dimensional vector in the low-dimensional vector sequence, determining a matching code element vector that matches each low-dimensional vector among a set of multiple code element vectors includes: For each low-dimensional vector in the low-dimensional vector sequence, calculating its similarity with each code element vector among the multiple code element vectors to obtain multiple similarities corresponding one by one to the multiple code element vectors; Determining the code element vector corresponding to the target similarity as the matching code element vector that matches; wherein, the target similarity is greater than other similarities among the multiple similarities.

3. The system state evolution prediction method for a complex system according to claim 1, characterized in that The step of, according to the energy landscape, determining a first matching code element vector corresponding to the prediction time in the code element topological structure includes: According to the energy landscape, determining a target matching code element vector corresponding to the first system state at the current time in the code element topological structure, and determining a first transfer path from the current time corresponding to the first system state to the prediction time in the code element topological structure; Taking the position of the target matching code element vector in the code element topological structure as a starting point, calculating a first state transfer probability of the target matching code element vector transferring to each matching code element vector in the code element topological structure along the first transfer path, to obtain multiple first state transfer probabilities corresponding one by one to the multiple matching code element vectors in the code element topological structure; Determining the matching code element vector corresponding to the target first state transfer probability as the first matching code element vector; wherein, the target first state transfer probability is greater than other state transfer probabilities among the multiple state transfer probabilities.

4. The method for predicting the system state evolution of a complex system according to claim 3, wherein After the step of determining the matching code element vector corresponding to the target first state transfer probability as the first matching code element vector, further including: Obtaining parameter values of at least one state control parameter; Predicting a second transfer path from the current time to the prediction time in the code element topological structure according to the energy landscape and the parameter values of the at least one state control parameter. Starting from the position of the target matching symbol vector in the symbol topological structure, calculate the second state transition probability of the target matching symbol vector transferring along the second transfer path to each matching symbol vector in the symbol topological structure, and obtain multiple second state transition probabilities corresponding one-to-one to multiple matching symbol vectors in the symbol topological structure; Determine the matching symbol vector corresponding to the target second state transition probability as the second matching symbol vector; wherein, the target second state transition probability is greater than other second state transition probabilities among the multiple second state transition probabilities; Perform high-dimensional expansion on the second matching symbol vector to obtain the second predicted system state corresponding to the prediction moment.

5. The method for predicting the system state evolution of a complex system according to claim 3, wherein Before obtaining the historical system state sequence of the complex system and obtaining the prediction moment, it further includes: Based on the graph neural network and the Fokker-Planck equation, establish a target state transition prediction model; According to the energy landscape, determine the target matching symbol vector corresponding to the first system state at the current moment in the symbol topological structure, and determine the first transfer path from the current moment corresponding to the first system state evolving to the prediction moment in the symbol topological structure, including: Input the energy landscape and the symbol topological structure into the target state transition prediction model, and obtain the target matching symbol vector and the first transfer path based on the target state transition prediction model.

6. The method for predicting the system state evolution of a complex system according to claim 5, wherein Before obtaining the historical system state sequence of the complex system and obtaining the prediction moment, it further includes: establish a target encoding model and establish a target decoding model coupled with the target encoding model; For each low-dimensional vector in the low-dimensional vector sequence, determine the matching symbol vector that matches each low-dimensional vector among the set of multiple symbol vectors, and determine the matching symbol vector sequence corresponding to the low-dimensional vector sequence based on the matching symbol vectors that match each low-dimensional vector, including: Input the multiple symbol vectors and the low-dimensional vector sequence into the target encoding model, and obtain the matching symbol vector sequence based on the target encoding model; The performing high-dimensional expansion on the first matching symbol vector to obtain the predicted system state corresponding to the prediction moment includes: Input the first matching symbol vector into the target decoding model, and obtain the predicted system state based on the target decoding model.

7. The method for predicting the system state evolution of a complex system according to claim 6, wherein The establishing the target encoding model and establishing the target decoding model coupled with the target encoding model includes: Jointly train the initial encoding model, the initial decoding model, and the initial state transition prediction model to obtain the target encoding model, the target decoding model, and the target state transition prediction model; wherein, the first loss function of the initial encoding model and the initial decoding model is the two-norm distance between the first input vector input into the initial encoding model and the second output vector output by the initial decoding model; the second loss function of the initial state transition prediction model is the two-norm distance between the second input vector and the second output vector of the initial state transition model.

8. A system state evolution prediction device for a complex system, characterized in that It includes: An acquisition module, configured to acquire a historical system state sequence of a complex system and acquire a prediction moment; A dimensionality reduction module, configured to perform dimensionality reduction processing on each historical system state in the historical system state sequence to obtain a low-dimensional vector sequence corresponding to the historical system state sequence; A determination module, configured to determine, for each low-dimensional vector in the low-dimensional vector sequence, a matching code element vector that matches each low-dimensional vector among a set of multiple code element vectors, and determine a matching code element vector sequence corresponding to the low-dimensional vector sequence based on the matching code element vectors that match each low-dimensional vector; A processing module, configured to construct a Voronoi diagram including all the matching code element vectors in the matching code element vector sequence based on the matching code element vector sequence, and construct a code element topological structure based on the Voronoi diagram; A calculation module, configured to obtain the distribution of each matching code element vector in the matching code element vector sequence based on the multiple historical system states and the code element topological structure, and calculate the energy landscape of the complex system based on the distribution; A prediction module, configured to determine a first matching code element vector corresponding to the prediction moment in the code element topological structure according to the energy landscape; A high-dimensional expansion module, configured to perform high-dimensional expansion on the first matching code element vector to obtain a predicted system state corresponding to the prediction moment.

9. An electronic device, comprising a memory, a processor, and a computer program stored on the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the system state evolution prediction method of the complex system according to any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the system state evolution prediction method of the complex system according to any one of claims 1 to 7.

Images