690 results about "Feedforward neural network" patented technology

The invention discloses a process parameter optimization control method and system for an automobile injection molding part, and belongs to the technical field of data processing, and the method comprises the steps: collecting sample data to train a feedforward neural network based on the historical processing record of a target injection molding part, and constructing a product quality prediction plug-in; performing parameter fluctuation influence analysis according to the historical processing record, and constructing a fluctuation influence analyzer; determining product demand characteristics of the target injection molded part in combination with the automobile use scene; and by taking the product demand characteristics as expectations, optimizing the process parameters by using the product quality prediction plug-in and the fluctuation influence analyzer, and outputting the optimal process parameters for processing control. The problems of large product quality fluctuation and low production efficiency caused by insufficient injection molding process parameter control precision in the prior art are solved. By combining historical records, product quality prediction plug-ins and fluctuation analysis, precise optimization of process parameters is realized, the product quality is stabilized, and the controllability and efficiency of the production process are improved.

The invention relates to the technical field of electric power engineering purchase demand prediction, in particular to an electric power engineering purchase demand prediction system based on machine learning, and the system comprises the steps: obtaining historical purchase data, construction progress information and electric power engineering design parameters, carrying out the standard stage division and time alignment, and constructing a stage sequence model reflecting the material use rhythm; and a coupling factor matrix is generated based on the material co-occurrence frequency and the stage position relationship, and the modeling capability of the model for the material cooperation relationship is enhanced. And the stage time sequence features, the coupling information and the structured engineering parameter vectors are fused and input into a regression prediction model, so that accurate mapping of material demands and multi-dimensional engineering features is realized, and the purchase prediction precision in a target period is improved. A deviation sequence is constructed based on historical prediction errors, and error correction is performed through a feedforward neural network, so that prediction accuracy and response capability are effectively improved, and resource waste and construction delay are reduced.

Predicting the salience of one or more data entities to a particular (target) data entity from among a plurality of data entities may comprise generating a graph of the plurality of data entities and a machine-learned model architecture that predicts the salience of the one or more data entities output by the machine-learned model architecture using the graph. For example, the machine-learned model architecture may comprise a first machine-learned model for generating an embedding using the content of the target data entity, a second machine-learned model for generating a vector using the data type indicated by the target data entity, and a third machine-learned model (e.g., a graph neural network or other feed-forward neural network) for generating a contextual representation of the target data entity to which other contextual representations associated with the plurality of data entities may be compared (e.g., using Euclidean distance, cosine similarity, dot product).

The invention discloses a large language model accelerator architecture based on a three-dimensional NAND flash memory, and belongs to the technical field of calculation, reckoning or counting. The architecture comprises a three-dimensional NAND flash memory used for executing feedforward neural network calculation; the auxiliary calculation unit is used for executing attention mechanism calculation; the DRAM chip is used for storing attention mechanism related weights and KV cache; and the interconnection resource is used for realizing data interaction among the components. Wherein the three-dimensional NAND flash memory comprises a logic chip and an NAND array chip, the logic chip is used for controlling and executing calculation, and the NAND array chip is used for storing weights and participating in calculation. The invention further provides a scheduling method based on KV cache awareness. The scheduling method comprises decomposition and dynamic allocation of calculation tasks. Through collaborative design of the hardware architecture and the scheduling method, the memory wall bottleneck in large language model calculation can be relieved, the energy consumption is reduced, and the overall calculation performance is improved.

The invention relates to a hybrid network 4D flight path prediction method based on LSTM and Transform, and belongs to the technical field of flight path prediction. Comprising the following steps: inputting preprocessed flight path data into an embedding layer, mapping the preprocessed flight path data into high-dimensional feature representation, and forming an input sequence; inputting the input sequence into the LSTM network to generate an output representation with a time-dependent feature; performing global feature extraction on the output representation of the LSTM network by using a multi-head attention mechanism so as to capture a dependency relationship and relevance between different trajectory data points; performing normalization processing on global features output by the multi-head attention mechanism, and further enhancing feature expression ability through a feedforward neural network; a mask multi-head attention mechanism is adopted in a decoder, and prediction features are output; and mapping the prediction features output by the decoder to a target space through a linear layer, and generating a final result of flight path prediction. The method is excellent in performance in complex flight path prediction tasks.

The invention discloses a remote sensing image segmentation method fusing frequency modulation and spatial perception, and the method comprises the steps: obtaining and preprocessing an original remote sensing image, and generating a standardized input image; the image is input into a multi-scale frequency domain enhanced feature extraction network, features are extracted step by step according to a plurality of feature levels, each level realizes frequency adaptive semantic enhancement through frequency domain modulation transformation and spatial feature fusion, and deep feature expression is enhanced through feedforward neural network modeling and residual connection output and cross-level residual fusion introduction; the final multi-scale features are decoded through a decoding module, the spatial resolution is recovered, and a pixel-level segmentation result is generated; and constructing a composite loss function containing classification errors, boundary perception and frequency consistency items, and carrying out optimization training on the network. According to the method, semantic complementarity of a remote sensing image in a frequency domain and a space domain is fully mined, so that segmentation precision and robustness of a ground object target in a complex scene are improved, and the method has good generalization ability and engineering practicability.

The invention provides a prediction method and device for operation parameters of a virtual power plant, a storage medium and a computer program product. Comprising the steps of performing data preprocessing on acquired influence factor data; using a time sequence decomposition algorithm to decompose the influence factor data into three-dimensional sequences including a trend component, a season component and a residual component, and segmenting each sequence into sequence blocks and mapping the sequence blocks into time sequence feature vectors; performing K-means clustering on word embedding used for pre-training the large language model, and selecting K clustering centers as semantic anchor points to be spliced with the time sequence feature vectors; finely adjusting the position embedding parameters of the large language model, the weight of the feedforward neural network in residual connection and the parameters of the normalization layer; and according to the semantic enhancement time sequence feature vector, predicting to obtain a trend component, a season component and a residual component of each dimension, and carrying out splicing and reverse normalization processing on the components to obtain a prediction result. According to the prediction method, the accuracy and real-time performance of virtual power plant load and electricity price prediction are improved.

The invention discloses an H-bridge key equipment service life and system reliability evaluation method and system for a cascade network construction type energy storage system, and belongs to the technical field of power system automation. The method comprises the following steps: firstly, extracting task profile parameters under multiple time scales, and constructing a time sequence feature model; secondly, estimating a hot spot temperature sequence of the IGBT device and the capacitor based on a multilayer feedforward neural network; then, in combination with a continuous extreme point paired temperature cycle extraction method and a Miner linear cumulative damage criterion, the damage factor and the residual life of the device are evaluated; then, task profile samples are expanded based on a generative adversarial network with gradient penalty, and life distribution and reliability indexes of key devices under different profiles are calculated; and finally, based on H-bridge series structure mapping device level information, constructing a system level reliability model, obtaining system failure rate, average fault-free operation time and a reliability function, and realizing health state perception and reliability quantitative evaluation of the energy storage system.

The embodiment of the invention discloses an inference method and device for a large language model, which are used for reducing bandwidth consumption of inference of the large language model. The method comprises the steps that a parallel reasoning system receives reasoning tasks of a large language model, wherein the reasoning tasks comprise a calculation-intensive task and a self-attention calculation task; the processor executes calculation-intensive tasks, generates intermediate data and sends the intermediate data to the near-memory calculation module, and the calculation-intensive tasks comprise one or more of the following tasks of the large language model: a feedforward neural network calculation task, a projection task and a layer normalization task. And the near memory calculation module executes a self-attention calculation task according to the intermediate data to generate a near memory calculation result. And the processor generates a reasoning result corresponding to the reasoning task based on the proximity calculation result.

The invention discloses an antibacterial peptide recognition method and system based on a sequence-structure dual-channel neural network, and the method comprises the steps: splicing amino acid features and amino acid-level manual features extracted by ProtT5 to obtain peptide embedding, and transmitting the peptide embedding to a sequence channel composed of a plurality of Transform blocks to extract the sequence features of the peptide; predicting a three-dimensional structure of the peptide by using ESM-Fold to construct an adjacency graph, taking amino acid features obtained by ESM-2 as node features of the adjacency graph, and performing layer-by-layer extraction and enhancement by fusing structural channels of multi-head graph attention, a residual network, layer normalization and a feedforward neural network; and carrying out maximum pooling and splicing on the sequence features and the structural features, and then, carrying out antibacterial peptide prediction. According to the method, a multi-feature fusion strategy is adopted, meanwhile, the three-dimensional structure information of the antibacterial peptide is introduced, and the sequence and the structural features are fused through a two-channel architecture, so that the recognition accuracy of the antibacterial peptide is effectively improved.

Aspects of the disclosure include methods for leveraging retrieval augmented generation (RAG) over a graph neural network (GNN) for edge building and the generation of reason-aware graph recommendations. A method can include constructing a graph neural network from an input graph having a plurality of nodes and one or more edges. The graph neural network includes one or more internal layers, each internal layer having one or more node vectors encoding a K-hop neighborhood for a target node of the plurality of nodes. RAG data including non-graph contextual data is retrieved for each of the plurality of nodes and transformed into embeddings using a large language model encoder. The RAG embeddings are encoded into node vectors of the graph neural network. The graph neural network generates a representation for the target node that is transformed by a feed forward neural network tower into an output vector.

The invention belongs to the technical field of target detection, and discloses an unmanned aerial vehicle image target detection method based on an improved RTDETR model, and the method introduces an FAPPA module in a backbone network, and effectively alleviates the information loss problem of small target features in a deep network. According to the method, a feedforward neural network of an AIFI module in an encoder is replaced by an EDFFN module, and high-frequency features such as edges and textures are selectively enhanced by using a patchwise FFT strategy and learnable frequency domain weights. Besides, according to the method disclosed by the invention, adaptive alignment of multi-scale features is realized by utilizing a Zoomcat module, and local details and global semantics are fused by utilizing a convolution attention double-path architecture of a CAFMFusion module, so that complementary information of different levels is fully utilized. According to the method, the unmanned aerial vehicle image target detection capability can be remarkably improved, and small target features in a complex background can be identified.

The invention discloses an energy storage rapid compensation method and system based on track overhead line system voltage fluctuation feature recognition, and relates to the technical field of track traffic traction power supply. The method comprises the following steps: synchronously acquiring voltage and current signals of the overhead line system, and extracting a composite feature vector containing disturbance root attributes and transient change rate; on the basis of the vector, a transient energy vacancy sequence covering the ultra-short-term future is output in real time through a prediction model fusing dynamic phasor analysis and a feedforward neural network; according to the sequence, a cooperative control strategy fusing overshoot and active damping is generated in combination with a proximity compensation principle and a voltage recovery state so as to drive an along-line energy storage unit; the strategy is executed in advance, compensation energy is injected, and finally double closed-loop correction is conducted based on the voltage residual error. According to the method, the problems of response lag, unclear disturbance identification, poor collaboration and lack of adaptive ability in the prior art are solved, and advanced, accurate, collaborative and self-optimized rapid suppression of the voltage fluctuation of the overhead line system is realized.

The invention discloses a multi-context semantic recognition and understanding method based on a large language model. The method comprises the following steps: S1, generating a semantic unit sequence; s2, constructing a context nested vector sequence; s3, constructing context priori representation, and generating a semantic representation sequence; s4, outputting a state vector of the semantic path by adopting a gating loop unit, obtaining a matching degree score according to a feedforward neural network, and determining a deliberate map tag and an alternative intention tag; s5, slot field extraction and semantic filling are completed, and a structured semantic task unit is generated; and S6, completing semantic recognition and service response closed loop. According to the method, by introducing a prefix regulation and control mechanism and a multi-context semantic modeling structure, the accuracy of intention recognition in multiple rounds of dialogues and the consistency of the context generated in response are remarkably improved, and the method is suitable for natural language understanding scenes of multi-language intelligent customer service, cross-context man-machine interaction and complex task driving.

The invention provides an electric power insulator defect detection method and system based on deep learning, and belongs to the technical field of electric power equipment insulation control detection. The method comprises the following steps: synchronously acquiring an original visible light image, an original infrared thermal imaging image and original ultrasonic data of a target electric power insulator; after preprocessing, data enhancement is carried out, then standard data is transmitted to a feature extraction module, and visible light features, infrared features and ultrasonic features are extracted; performing feature splicing, mapping to a shared space, inputting into a multi-head attention mechanism, generating a combined parallel result, and based on the combined parallel result, generating a fusion feature vector from the visible light feature, the infrared feature and the ultrasonic feature through a feedforward neural network; and outputting the defect type of the target power insulator through an output layer of the cross-modal image matching network. The method provided by the invention has higher precision and stronger robustness.

The invention provides an electro-hydraulic servo pump control method and system for controlling hydrogen compression based on fuzzy nerve, and relates to the technical field of hydrogen compression, and the method comprises the steps that the current state of a hydraulic cylinder controlled by an electro-hydraulic servo pump is determined, and an input control variable is obtained; determining an output control variable, establishing a fuzzy control rule, and dividing a basic discourse domain; performing composite defuzzification, and obtaining a corrected weight coefficient through a feedforward neural network; the discrete rule amplitude correction factor and the weight coefficient correction are dynamically optimized and corrected; neural network training is established, a loss function evaluation model is utilized, and hydrogen compression control is achieved through a double-buffering mechanism. Fuzzy PID control is adopted, composite defuzzification is carried out according to the real-time state change of a system in combination with fuzzy logic, a membership function output by the fuzzy PID is divided according to regions, a control signal of a driver is synthesized, and hydrogen compression control is achieved; system output is adjusted according to different working conditions, power loss is reduced, and system control precision is improved.

The invention provides an optical element surface type simulation method based on a physical information neural network, and relates to the technical field of element design and simulation, and the method comprises the steps: constructing a feedforward neural network; generating a deformation control equation model library; constructing a three-dimensional simulation database; forming a machine learning parameter library; training the PINN deformation simulation model in combination with a deformation control equation model library, a three-dimensional simulation database and a machine learning parameter library, and screening out a target PINN deformation simulation model meeting a preset condition; inputting the design parameters, the support structure parameters and the given external load parameters of different combinations into the target PINN model to obtain a key performance prediction result of the optical element; according to the method, the key performance of the element under various parameter combinations can be rapidly evaluated, the design iteration period is shortened, the trial and error cost is reduced, direct data support is provided for supporting structure optimization design, and the bottleneck problem that a traditional method restricts the research and development efficiency of a high-precision optical system is effectively solved.

The invention provides a medical image segmentation method based on a B-spline function and a Transform, and solves the technical problem that in the Transform, the parameter quantity of the MLP is huge, and the MLP generally does not have interpretability, so that the generalization ability and the characterization ability of a model are weak. According to the technical scheme, the method comprises the following steps of 1, preprocessing an input medical image; step 2, replacing a self-attention mechanism in Transfomer and a weight parameter of an MLP in a feedforward neural network by using a B-spline-based spline function; step 3, inputting the preprocessed image data set into an encoder and decoder network based on a B-spline function for training to obtain an optimal model; and step 4, after training is completed, inputting the medical image verification set into the obtained optimal model. According to the method, the focus can be segmented more accurately, and the accuracy and robustness are improved.

The invention belongs to the technical field of remote sensing image analysis, and provides a multi-source remote sensing building extraction method based on hybrid experts, electronic equipment and a storage medium. The method comprises the steps of basic model construction, two-way structure model construction, optical feature extraction, earth surface feature extraction, two-way feature fusion, ViT coding, feature injection and extraction and semantic alignment step-by-step up-sampling. According to the invention, a dual-path structure is adopted to assist the multi-scale encoder and the bidirectional attention fusion module, and DSM data and optical images are used for feature fusion, so that the capability of distinguishing buildings from backgrounds is improved; the hybrid expert LoRA structure is introduced into the ViT encoder, parameters of the feedforward neural network are dynamically adjusted, the adaptability and flexibility of the model to input features are further enhanced, and the calculation complexity is reduced.

The invention relates to the technical field of channel environment monitoring, and discloses a channel environment safety monitoring and early warning method and system based on multi-sensor fusion, which align visible light video, infrared thermal imaging, gas and vibration sensor data through space-time reference to ensure isomerous data synchronism and space consistency. Modal exclusive time sequence feature extraction is carried out, and unique space-time mode information is mined. A context sensing mechanism is introduced, key information such as an operation and maintenance plan is intelligently extracted in combination with a current environment state, dynamic weighting and deep fusion are performed on multi-sensor time sequence features through an attention fusion mechanism, and the importance of different sensor features is adaptively adjusted. And finally, inputting the fused channel environment state into a feedforward neural network model, and outputting an event probability and a risk score. Therefore, safety monitoring accuracy and reliability are remarkably improved, false alarms are reduced, potential safety hazards are recognized in advance, quantitative decision basis is provided, and continuous and safe operation of a channel environment is guaranteed.

The invention relates to a traffic flow prediction method based on temporal-spatial feature fusion Transform in an edge environment, and the method comprises the steps: carrying out the feature embedding and coding of original traffic flow data; then, constructing a multi-head convolution low-rank decomposition attention mechanism to capture a long-term time dependency relationship and obtain local context information; then, constructing an attention map convolution to capture a spatial dependency relationship; and finally, performing adaptive fusion on the spatial-temporal characteristics through a gating unit, and further realizing accurate prediction of future traffic flow by using a feedforward neural network and a linear layer. The method can improve the prediction precision, effectively reduce the resource overhead, and improve the prediction efficiency.

The invention discloses a flow field prediction-oriented full-process automatic joint reduced-order modeling method, which comprises the following steps of: specifying a target physical field parameter space, and randomly generating a sample space according to a Latin hypercube sampling method; constructing a full-process automatic simulation tool chain, driving target physical field numerical calculation and generating a training data set; carrying out singular value decomposition-based intrinsic orthogonal decomposition on the output physical field data, and only retaining first r main feature components to construct a reduced-order data set; constructing a multi-input multi-output full-connection feedforward neural network, and modeling and training a nonlinear mapping relation between input parameters and reduced-order features; new working condition parameters are input, reduced-order features are predicted through the trained neural network, distribution of a target physical field is reconstructed according to a singular value decomposition reduction matrix, and more flexible and reliable technical support is provided for reducing the training cost of a reduced-order model and improving simulation efficiency.

Disclosed in embodiments of the present application are a reasoning method and device for a large language model, which are used for reducing the bandwidth consumption of large language model reasoning. The method of the embodiments of the present application comprises: a parallel reasoning system receives a reasoning task of a large language model, wherein the reasoning task comprises a compute-intensive task and a self-attention computing task; a processor executes the compute-intensive task, generates intermediate data, and sends the intermediate data to a near-memory computing module, wherein the compute-intensive task comprises one or more tasks of the large language model: a feedforward neural network computing task, a projection task, and a layer normalization task; the near-memory computing module executes the self-attention computing task on the basis of the intermediate data to generate a near-memory computing result; and on the basis of the near-memory computing result, the processor generates a reasoning result corresponding to the reasoning task.

The invention relates to a power battery performance degradation prediction method and system based on an electrochemical-thermal-mechanical-neural network model, and belongs to the technical field of battery management. According to the method, an enhanced single particle model considering liquid phase diffusion, an integrated thermal model, a particle mechanical stress model caused by concentration gradient and an aging mechanism model considering SEI growth, lithium precipitation and active material loss are coupled to establish ETMD, and based on the ETMD, the identification efficiency of aging model parameters is low aiming at the problem that a PSO algorithm is low. Under a circulating aging experiment of a limited path, the characteristic parameters of the aging model are obtained by using the feedforward neural network, and the parameters are substituted into the ETMD model again to predict the capacity, the power attenuation curve and the corresponding main attenuation mechanism under the current path, so that the prediction precision and the interpretability of the performance attenuation of the battery under the full path are improved. According to the method, the modeling capability of battery performance degradation under different working conditions can be effectively improved, and a theoretical basis and technical support are provided for residual life prediction and safety management of the power battery.

The invention relates to an inertial sensor sensitive structure design method based on a neural network interpolation algorithm, and the method comprises the steps: defining geometric parameters and material parameters of an inertial sensor sensitive structure, and generating a three-dimensional geometric configuration related to the inertial sensor sensitive structure; performing finite element simulation on the three-dimensional geometric configuration to obtain a simulation result containing performance indexes; adjusting the geometric parameters and the material parameters for a plurality of times, generating a large number of simulation results, forming a simulation database, and forming a sample data set formed by input parameter combination and target performance pairing according to the simulation database; a feedforward neural network is adopted, the sample data set is input into the feedforward neural network for training, an interpolation prediction model is obtained, and the interpolation prediction model is adopted for prediction; through an innovative framework of simulation data driving, neural network modeling and closed-loop optimization, a systematic solution is provided for sensor development which is high in performance, low in cost and rapid in iteration.

The invention relates to the technical field of cable monitoring, in particular to a graphene cable monitoring system based on deep learning, which is used for extracting time-frequency domain characteristics by acquiring current, voltage, electromagnetic wave signals and temperature distribution data and utilizing Fourier transform and wavelet transform to improve the data identification degree. Gaussian filtering noise reduction is carried out on temperature data, and an abnormal hot spot area is identified through image segmentation, so that the local overheating detection capability is improved. And the edge calculation module fuses multi-source data, eliminates acquisition delay by utilizing feature alignment, and improves data synchronism and fusion quality. Through a multi-head self-attention mechanism, time sequence characteristics of historical monitoring data are extracted, and change modes of current, voltage, electromagnetic wave and temperature distribution are learned. And calculating an attention weight matrix to extract correlation between time steps, and forming a time sequence feature matrix. The characteristic matrix is subjected to nonlinear transformation through a feedforward neural network, cable state parameters of a future time step are predicted, the cable state is evaluated in advance, and the fault risk is reduced.

The invention relates to a bearing fault diagnosis method based on variational mode decomposition and time sequence partitioning cross attention fusion, which comprises the following steps: acquiring an original vibration acceleration signal of a rolling bearing, and constructing a standardized original data set; segmenting the standardized original data into a plurality of data blocks, and generating a time domain embedding feature; based on the time domain embedded features, extracting high-order global time domain features by using a multi-head self-attention mechanism, residual connection and a feedforward neural network; decomposing the standardized original data into a plurality of intrinsic mode functions, and extracting frequency domain distribution features through a convolutional neural network; taking the frequency domain distribution characteristics as query vectors, retrieving and matching related fault context information in the global time sequence characteristics, realizing weighted fusion of time-frequency modes, inputting fused fault representation vectors into a classifier, and calculating a result of a bearing health state; and constructing a loss function containing label smoothing and a dynamic learning rate scheduling strategy, and carrying out iterative optimization on model parameters until the model converges.

The invention discloses an electro-hydraulic servo system finite time accurate tracking control method and system. The method specifically comprises the steps that a mathematical model of a single-rod electro-hydraulic position servo system is established; based on a multilayer feedforward neural network, constructing a finite time neural network extended state observer, and estimating matching and non-matching unknown function disturbances and system states suffered by the single-rod electro-hydraulic servo system; constructing a finite time accurate tracking control algorithm of the single-rod electro-hydraulic servo system for model uncertainty compensation, and designing a finite time multilayer neural network adaptive law with a fractional order; and an initial value of a neural network weight parameter, an adaptive law matrix and a controller parameter are selected to realize the compensation of the uncertainty of a system model, so that the output of the system tracks an expected control target. According to the method, the steady-state tracking error of the single-rod electro-hydraulic servo system controller is reduced in finite time, the uncertainty compensation capability of the model is improved, and the tracking precision and the tracking speed of the single-rod electro-hydraulic servo system are improved.

The invention discloses an electric power system large-scale language model tuning method, system and device based on LoRA technology optimization and a medium, and belongs to the technical field of electric power system large-scale language models, and the method comprises the steps: obtaining historical language data formed in the operation process of an electric power system, and converting the historical language data into a training sample; constructing a task data set based on the training sample; constructing a pre-training language model; receiving a to-be-processed input text, extracting semantic features of the input text, and generating an output sequence corresponding to the input text. According to the method, the low-rank weight path is inserted in the feed-forward neural network in a parallel form, and an independent path is bound for each type of tasks, so that differential fine tuning of the electric power semantic tasks is realized on the premise of not modifying the structure and parameters of an original model, the overall tuning cost of the model is reduced, and the stability of the original model is kept.

The invention discloses a flying capacitor voltage prediction method based on a neural network. The method comprises the following steps: S10, collecting a switching signal, an output voltage and a capacitor voltage true value of a flying capacitor multi-level converter FCMC in a no-load to full-load operation process through a model prediction controller MPC; s20, a double-layer feed-forward neural network is constructed, a defined input vector and a defined output vector are recorded, the input vector comprises a switching signal and an output voltage at the current moment and a flying capacitor voltage estimated value at the previous moment, and the output vector is the flying capacitor voltage estimated value at the current moment; s30, repeating the step S10 and the step S20 by taking different capacitor voltages as initial conditions; s40, performing off-line training on the neural network by using a back propagation algorithm BP, optimizing network weight and bias, and enabling network output to approach real voltage of flying capacitor operation at the moment; and S50, deploying the trained neural network to an FPGA controller, and estimating the flying capacitor voltage online in real time in combination with a model predictive control MPC algorithm.