Transformer state prediction method and system based on hybrid neural network

Abstract

The application discloses a transformer state prediction method and system based on a hybrid neural network, and belongs to the technical field of intelligent operation and maintenance of a power system and time series prediction. In view of the problem that existing prediction methods are difficult to simultaneously consider global physical coupling modeling and long-range nonlinear evolution robust fitting among multiple variables, the application decomposes an original sequence into independent frequency bands by discrete wavelet transform, accurately reconstructs multiple variable physical coupling by a quantum channel entanglement module, extracts long-range fluctuation characteristics by means of a residual quantum Token hybrid module and guarantees training stability, and finally outputs a high-precision result. The application realizes accurate long-time prediction of a future state of a transformer while keeping training robustness, and provides a reliable decision basis for advanced warning and load scheduling of a power grid dispatching center.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance and time series prediction technology for power systems, and in particular to a transformer condition prediction method and system based on hybrid neural networks. Background Technology

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

[0003] In smart grid dispatching, power transformers, as core hub equipment, require accurate long-term prediction of oil temperature and load for preventing overheating faults, ensuring power supply reliability, and optimizing load dispatching. However, in actual operation, transformer monitoring data exhibits strong non-stationarity (such as distribution shifts between morning and evening peak and off-peak periods, and seasonal cyclical changes) and complex physical coupling characteristics among multiple variables (thermodynamic and electrical nonlinear interactions exist between oil temperature, current, voltage, and load).

[0004] While existing time series forecasting methods each have their own focus, they generally face a dual dilemma: on the one hand, classical models, represented by channel-independent strategies, while highly efficient in training, fundamentally sever the global dependencies between variables; on the other hand, traditional deep networks, when modeling long sequences, either struggle to converge due to vanishing gradients or sacrifice parameter efficiency and inference speed due to excessive network stacking. Furthermore, even with the introduction of quantum machine learning, existing hybrid quantum-classical architectures still face bottlenecks in long-term forecasting tasks, including training instability, susceptibility to quantum noise, and limited optimization capabilities.

[0005] Therefore, existing prediction methods cannot simultaneously take into account both accurate modeling of global physical coupling among multiple variables and robust fitting of long-sequence nonlinear evolution. As a result, in the actual operation of transformers, when faced with complex operating conditions such as sudden increases in load during morning and evening peak hours and sudden disturbances caused by extreme weather, the models either deviate due to the severing of physical correlations or fail to predict due to gradient instability. They cannot provide the dispatch center with advanced and reliable early warning of oil temperature exceeding the threshold and a basis for load dispatching decisions. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a transformer state prediction method and system based on hybrid neural networks. By using multi-scale wavelet decomposition and hybrid quantum-classical neural network collaborative modeling, it achieves high-precision long-term prediction of transformer oil temperature and load, providing advanced early warning for power grid dispatching.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solution: In a first aspect, the present invention provides a transformer state prediction method based on a hybrid neural network, comprising: Obtain the raw multivariate time series data of transformer operation; The transformer is decomposed into a low-frequency sub-band reflecting the long-term evolution trend of the transformer and multiple high-frequency sub-bands reflecting sudden disturbances through discrete wavelet transform. Each sub-band is then input into a hybrid quantum-classical neural network, where the quantum channel entanglement module of the hybrid quantum-classical neural network recalibrates the global physical coupling relationship between multiple variables. The recalibrated sub-band features are sliced ​​and linearly embedded to obtain patch embedding features; the patch embedding features are input into the residual quantum token hybrid module in the hybrid quantum-classical neural network to extract cross-time-period dependencies and obtain the enhanced temporal representation; The enhanced time series representation is mapped to the future prediction values ​​of each sub-band through a linear prediction head, and the transformer state prediction result is obtained after reconstruction.

[0008] Furthermore, the original multivariate time series data includes at least transformer oil temperature, three-phase current, three-phase voltage, and active load.

[0009] Furthermore, the quantum channel entanglement module aggregates time-dimensional information using global average pooling; After being compressed to the qubit dimension by linear projection, the physical coupling relationship between the multivariable channels is modeled by the strong entanglement layer and the two-qubit entanglement gate in the parameterized quantum circuit; The original features are recalibrated using channel attention weights.

[0010] Furthermore, the quantum channel entanglement module divides the recalibrated features into discrete patch sequences according to the time span. Each patch contains multiple consecutive time steps, which are then mapped to patch embedded features through a learnable linear embedding layer.

[0011] Furthermore, the residual quantum token mixing module adopts a dual-stream architecture that processes classical linear mapping and quantum nonlinear mapping in parallel and fuses residuals; Among them, classical linear mapping captures the macroscopic trend of the sequence and serves as a stability anchor, while quantum nonlinear mapping extracts higher-order nonlinear fluctuations between patches through variable quantum circuits.

[0012] Furthermore, the quantum nonlinear mapping encodes the patch embedding features into quantum rotation gate parameters via a learnable projection layer; After angular embedding and multi-layer strong entanglement evolution, quantum features are extracted using Pauli-Z basis measurement, and then a nonlinear correction term is output through linear mapping. This term is then added to the output residual of the classical linear mapping and normalized.

[0013] Furthermore, the reconstruction includes sequentially performing inverse discrete wavelet transform to fuse the predicted values ​​of each sub-band, and performing inverse reversible instance normalization to recover the original physical dimensions of the prediction results, thereby obtaining the transformer state prediction results.

[0014] A second aspect of the present invention provides a transformer state prediction system based on a hybrid neural network, comprising: Data acquisition module: used to acquire raw multivariate time series data of transformer operation; Multi-scale decomposition and quantum channel recalibration module: used to decompose the transformer into low-frequency sub-bands reflecting the long-term evolution trend of the transformer and multiple high-frequency sub-bands reflecting sudden disturbances through discrete wavelet transform; each sub-band is input into a hybrid quantum-classical neural network, and the quantum channel entanglement module in the hybrid quantum-classical neural network recalibrates the global physical coupling relationship between multiple variables; The residual quantum temporal hybridization module is used to slice and linearly embed the recalibrated sub-band features to obtain patch embedded features; the patch embedded features are input into the residual quantum token hybridization module in the hybrid quantum-classical neural network to extract cross-time-period dependencies and obtain the enhanced temporal representation. Prediction output module: used to map the enhanced time series representation to the future prediction values ​​of each sub-band through the linear prediction head, and obtain the transformer state prediction results after reconstruction.

[0015] In a third aspect, the present invention also provides a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of the method as described in the first aspect.

[0016] In a fourth aspect, the present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method as described in the first aspect.

[0017] Compared with the prior art, the beneficial effects of the present invention are: This invention collects multi-dimensional data such as oil temperature, current, voltage, and load at a fixed frequency using sensors deployed on transformers. First, discrete wavelet transform is used to decompose the original sequence into low-frequency trend sub-bands reflecting daily cycles and seasonal evolution, and high-frequency disturbance sub-bands reflecting thunderstorms and sudden power outages, enabling independent and accurate feature extraction for each frequency band. Then, each sub-band is input into a hybrid quantum-classical neural network. A quantum channel entanglement module utilizes the global correlation characteristics of quantum entanglement gates to accurately reconstruct the thermodynamic and electrical physical coupling relationships between oil temperature, current, voltage, and load with extremely low parameter complexity. Next, the recalibrated sub-band feature slices are embedded as patch sequences. A residual quantum token hybrid module uses a classical linear path to ensure gradient stability and a quantum nonlinear path to capture complex fluctuations under extreme conditions. The residual fusion of these two modules effectively extracts long-range dependencies across time periods. Finally, through prediction head mapping and inverse reconstruction, a high-precision prediction of the transformer's future state is output. This provides a reliable decision-making basis for the dispatch center to identify oil temperature exceeding threshold risks, execute load transfers, and activate standby transformers in advance, significantly improving the safety and foresight of power grid operation. Attached Figure Description

[0018] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0019] Figure 1 This is a flowchart of the method according to Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the overall architecture of the model of the present invention; Figure 3 This is a schematic diagram of the variable quantum circuit (VQC) structure in the QCE of the present invention; Figure 4 This is a bar chart comparing the mean squared error (MSE) of ablation experiments of this invention on the ETTh1, ETTm1, and Weather datasets. Figure 5 This is a bar chart comparing the MAE (mean absolute error) of ablation experiments on the ETTh1, ETTm1, and Weather datasets of this invention. Figure 6 For the present invention in the ETTh1 dataset ( Noise robustness analysis plot on ) Figure 7 This is a heatmap of the Pearson correlation coefficients among the original meteorological indicators on the Weather dataset, based on the present invention. Figure 8 This is a schematic diagram of the quantum entanglement weight matrix learned by the QCE module of this invention; Figure 9This is a comparison of the model training convergence dynamics of this invention on the ETTh1 dataset; Figure 10 This is a comparison of the model training convergence dynamics of this invention on the ETTm1 dataset; Figure 11 This is a comparison diagram of the model training convergence dynamics of this invention on the Weather dataset. Detailed Implementation

[0020] The following detailed description is exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0021] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0022] In this embodiment of the invention, the data collection and processing strictly adhere to the requirements of relevant laws and regulations, obtaining informed consent or separate consent from the data subject, and conducting subsequent data use and processing within the scope of laws, regulations, and the data subject's authorization. All data acquisition in this embodiment is based on compliance with laws and regulations and user consent, representing the lawful application of the data.

[0023] Example 1 Compared to the shortcomings of existing technologies, for example, in actual power grid operation, transformer oil temperature and load are affected by a variety of factors such as morning and evening peak hours, seasonal changes, and sudden power outages, exhibiting strong non-stationarity and complex multivariate physical coupling characteristics. Existing prediction methods struggle to simultaneously consider both global physical coupling modeling and robust fitting of long-range nonlinear evolution, resulting in insufficient prediction accuracy and poor training stability.

[0024] To address the aforementioned issues, this invention first decomposes the original sequence into independent frequency bands using discrete wavelet transform to reduce mutual interference between different modes. Then, the quantum channel entanglement module accurately reconstructs the global physical dependencies between oil temperature, current, voltage, and load with linear parameter complexity. Simultaneously, the residual quantum token hybrid module ensures gradient stability with classical linear anchor points and accurately extracts long-range fluctuation features with quantum nonlinear paths. The synergy of these two approaches enables the model to accurately predict the future state of the transformer while maintaining training robustness, providing a reliable basis for early warning and load scheduling in the dispatch center.

[0025] In a typical embodiment of the present invention, such as Figure 1 As shown, a transformer state prediction method based on a hybrid neural network is disclosed, and the specific steps include: S1: Obtain the raw multivariate time series data of transformer operation; S2: The discrete wavelet transform decomposes the transformer into a low-frequency sub-band reflecting the long-term evolution trend and multiple high-frequency sub-bands reflecting sudden disturbances; each sub-band is input into a hybrid quantum-classical neural network, and the quantum channel entanglement module in the hybrid quantum-classical neural network recalibrates the global physical coupling relationship between multiple variables. S3: Slice and linearly embed the recalibrated subband features to obtain patch embedding features; input the patch embedding features into the residual quantum token hybrid module in the hybrid quantum-classical neural network to extract cross-time-period dependencies and obtain the enhanced temporal representation; S4: The enhanced time series representation is mapped to the future prediction values ​​of each sub-band through a linear prediction head, and the transformer state prediction result is obtained after reconstruction.

[0026] The above-mentioned transformer state prediction method based on hybrid neural networks will be described in detail below with reference to specific implementation methods.

[0027] In step S1, oil temperature, current, voltage, and load sensors deployed on the transformer are used to collect data in real time at a fixed sampling frequency, generating raw multi-dimensional time series data containing multiple physical dimensions. The raw multi-dimensional time series data includes at least transformer oil temperature, three-phase current, three-phase voltage, and active load.

[0028] In power grid dispatching scenarios, transformer condition prediction systems collect operational data in real time using a set of IoT sensors deployed on the transformer itself and transmission lines. Specifically, oil temperature sensors are installed inside the transformer tank to monitor changes in the top-level oil temperature. Current transformers and voltage transformers are installed at the high-voltage and low-voltage outgoing terminals, respectively, to obtain three-phase current and three-phase voltage signals. Simultaneously, power transmitters are installed on the bus side to measure active load. Each sensor collects data synchronously at a sampling frequency of once every 15 minutes. The data is then aggregated through field data acquisition terminals and transmitted via fiber optic or 4G wireless networks to the dispatch center's data server, forming continuous time-series data including oil temperature, three-phase current, three-phase voltage, and active load, which is then processed by subsequent prediction models.

[0029] In the monitoring of transformer operation in smart grids, multivariate time series data exhibit typical non-stationarity and complex dependencies. Non-stationarity manifests in the continuous changes in transformer oil temperature and load over time. This includes not only periodic fluctuations due to day-night cycles and long-term evolutionary trends caused by seasonal changes, but also dramatic spikes and abrupt changes caused by random events such as sudden high-power consumption, extreme weather, or equipment switching. These factors lead to significant shifts in data distribution across different time periods, and the statistical characteristics (such as mean and variance) are not constant. Simultaneously, the series contains complex dependencies: in the time dimension, the current oil temperature depends not only on the temperature value of the previous moment but also on the historical state over the past few hours or even days, exhibiting a long-range memory effect. In the variable dimension, there is a strong physical coupling between thermodynamics and electricity among oil temperature, three-phase current, three-phase voltage, and active load. For example, a sudden increase in load directly leads to an increase in current and oil temperature, while three-phase imbalance, in turn, affects voltage stability. This temporal nonstationarity, coupled with the multidimensional coupling between variables, makes it difficult for classical linear models and traditional deep neural networks to simultaneously capture long-range dependencies and reconstruct global multivariate correlations, thus posing a significant challenge to the accurate long-term prediction of the future state of transformers.

[0030] In this embodiment, in order to address the non-stationarity and complex dependencies in multivariate time series (for example, in the transformer prediction scenario, multivariate time series specifically refers to multiple physical quantity data that change continuously over time, which are synchronously collected by multiple types of sensors deployed on the transformer at a unified sampling frequency, including transformer top oil temperature, A / B / C three-phase current, A / B / C three-phase voltage, and active load, forming multiple-dimensional observation vectors at each moment, and arranged in chronological order to form a multivariate joint observation sequence), this invention proposes a hybrid quantum-classical neural network model (Quantum Wavelet Mixer, Q-WaveMixer) based on multi-scale wavelet decomposition.

[0031] like Figure 2 As shown, the Q-WaveMixer backbone adopts a multiresolution analysis (MRA) framework, and its main processing flow includes the following key stages: 1) Distribution Alignment and Decomposition: Given an input sequence First, Reversible Instance Normalization (RevIN) is used to mitigate the distribution shift problem in non-stationary data. Then, Discrete Wavelet Transform (DWT) is used to decompose the normalized sequence into a series of sub-band components of different frequencies (including low-frequency components that capture long-term trends and multiple high-frequency detail components that capture short-term fluctuations).

[0032] 2) Quantum Enhancement, Patching, and Embedding: For each decoupled frequency component, the model first reconstructs the physical dependencies between channels on a global time scale using a Quantum Channel Entangler (QCE). Then, a patching operation is performed, dividing the continuous time series into locally continuous subsequence segments (i.e., patches, or, the basic processing units tokens obtained after segmentation) by dividing the series into fixed time windows. Next, a learnable linear embedding layer maps these discrete patches to a high-dimensional latent space, generating feature vectors containing rich temporal context information for subsequent processing.

[0033] 3) Residual Mixing and Prediction: The generated patch embedding features are fed into a dual residual mixer, which consists of cascaded temporal mixing layers and positional feedforward layers. First, the features undergo batch normalization and dimension transposition, then are fed into a residual quantum token mixer (RQTM) to capture inter-patch nonlinear temporal dependencies. The output of the RQTM is again subjected to dimension reduction and normalization, and then fed into a feedforward neural network (FNN). The FNN operates on the hidden dimension of each patch, aiming to perform intra-patch feature integration to enhance the expressive power of each semantic unit. The entire process employs a residual connection mechanism, directly summing the inputs to the output of the FNN to promote gradient propagation in deep networks. Finally, the fully interacted features are mapped to the future predicted values ​​of the corresponding frequency components through a linear prediction head.

[0034] 4) Multi-scale Reconstruction: Finally, the prediction results of all frequency components are fused at multiple scales using the Inverse Discrete Wavelet Transform (IDWT), and the statistical distribution of the data is restored through the inverse denormalization of Reversible Instance Normalization (RevIN) to obtain the final prediction output. .

[0035] To efficiently capture inter-channel and inter-token dependencies within each frequency band, this invention deeply integrates classical deep learning modules with Variational Quantum Circuits (VQC), designing two core components: a Quantum Channel Entangler (QCE) and a Residual Quantum Token Mixer (RQTM). Specifically, the QCE utilizes quantum entanglement as an inductive bias to reconstruct higher-order physical couplings between variables at a lower parameter cost; while the RQTM leverages the high-dimensional nonlinear mapping capabilities of quantum circuits to capture complex temporal evolution and ensures gradient stability for long-sequence modeling through classical residual paths. Furthermore, to address the challenge of co-optimization among heterogeneous modules, this invention designs a Physics-Aware Quantum Hybrid Loss (QHL) during the training phase. By imposing physical consistency constraints, it guides model convergence, ultimately achieving an organic unity of quantum advantage and classical robustness.

[0036] In S2, the quantum channel entanglement module aggregates time-dimensional information using global average pooling; after being compressed to the quantum bit dimension by linear projection, the physical coupling relationship between multivariable channels is modeled by the strong entanglement layer and the two-qubit entanglement gate in the parameterized quantum circuit; and the original features are recalibrated using channel attention weights.

[0037] Specifically, in transformer condition prediction, the original oil temperature and load data often superimpose two different dynamic modes: long-term trends and sudden disturbances. If a single model is used to fit both simultaneously, the low-frequency trend may be submerged by high-frequency noise, and high-frequency abrupt changes may be dulled by long-term smoothing, making it difficult to maintain prediction accuracy. Therefore, this invention first performs discrete wavelet transform on the normalized time-series data, explicitly decomposing it into a low-frequency subband reflecting the long-term evolution trend of the transformer (such as slow rises caused by daily or seasonal cycles) and multiple high-frequency subbands reflecting sudden disturbances (such as spike fluctuations caused by thunderstorms or high-power equipment switching). This multi-resolution decomposition allows subsequent models to process data by frequency band, extracting features for different frequency bands without interference. However, regardless of the low-frequency or high-frequency components, there is a strong thermodynamic and electrical physical coupling between their various physical variables (oil temperature, current, voltage, load): for example, a sudden increase in load will simultaneously increase both current and oil temperature, while oil temperature changes lag behind load changes. To obtain this cross-band, cross-variable global coupling relationship, this invention inputs each subband into the quantum channel entanglement module in a hybrid quantum-classical neural network. By utilizing the non-local correlation characteristics of the quantum entanglement gate, the implicit dependencies between physical variables are globally recalibrated with extremely low parameter complexity. This enables the model to adaptively focus on the combination of key variables that is most dominant in the prediction at the current moment, thereby significantly improving the long-term prediction accuracy and robustness of the transformer's future state.

[0038] As an example, to overcome the parameter efficiency bottleneck of traditional classical layers (such as fully connected or convolutional layers) in capturing nonlinear high-order correlations between channels in multivariate time series, this invention proposes a Quantum Channel Entangler (QCE) module. This module is embedded in the multi-resolution branch after wavelet decomposition and utilizes the quantum entanglement properties in Hilbert space to globally recalibrate channel features.

[0039] Classical Squeeze and Mapping, as follows: Let the input feature tensor at a certain resolution be... First, in order to aggregate global information in the time dimension, this invention applies Global Average Pooling (GAP) to generate channel descriptors. :

[0040] Considering the qubit number limitation of noisy intermediate-scale quantum (NISQ) devices, this invention utilizes a classical linear projection layer. Compressing high-dimensional channel features to the quantum hidden space dimension (i.e., the number of qubits), and uses hyperbolic tangent (Tanh) activation to fit the range of quantum angle embeddings:

[0041] in, It is a learnable compression matrix.

[0042] Variational quantum entanglement evolution is as follows: Compressed feature vectors It is fed into a parameterized quantum circuit (PQC), such as Figure 3 As shown. The process consists of three stages: 1) Angle Embedding: Utilizing The rotating door encodes classical data into quantum states. Initial state Evolved into:

[0043] 2) Strongly Entangling Layers: To simulate the complex coupling between channels, a method is adopted... Strongly entangled layers. Each layer consists of a universal rotation gate acting on a single qubit. It consists of a two-qubit entanglement gate (CNOT) that acts on adjacent qubits:

[0044] In this process, the CNOT gate entangles different qubits, thereby implicitly reconstructing the higher-order interactions between the original channel features in the quantum state. This entanglement capability provides the model with a stronger feature mixing inductive bias than classical neural networks.

[0045] 3) Quantum Measurement: Measuring the expected value of each qubit under the Pauli-Z basis, decoding the quantum state back into a classical vector. :

[0046] Then, the original features are recalibrated using channel attention weights; feature recalibration is as follows: Map the quantum output back to the original channel dimension and generate channel attention weights. Weighting the original features:

[0047] In this way, QCE enables the model to adaptively focus on key frequency bands and variables by leveraging quantum parallelism.

[0048] In step S3, the quantum channel entanglement module divides the recalibrated features into discrete patch sequences according to the time span. Each patch contains multiple consecutive time steps, which are then mapped to patch embedding features through a learnable linear embedding layer. The patch embedding features are input into the residual quantum token hybrid module in the hybrid quantum-classical neural network to extract cross-time-segment dependencies and obtain the enhanced time-series representation. The residual quantum token hybrid module adopts a dual-stream architecture that processes classical linear mapping and quantum nonlinear mapping in parallel and fuses residuals. Among them, classical linear mapping captures the macroscopic trend of the sequence and serves as a stability anchor, while quantum nonlinear mapping extracts high-order nonlinear fluctuations between patches through variable quantum circuits.

[0049] Specifically, although the frequency sub-band features recalibrated by the quantum channel entanglement module already possess good inter-variable channel correlations, long-range dependencies across time periods still need to be modeled. Traditional methods typically model each time point directly, but adjacent time points in long sequences are often highly similar. Point-by-point processing is not only computationally redundant but also fails to capture macroscopic evolutionary patterns spanning hours or even days. To address this, this invention divides the recalibrated sub-band features along the time axis into multiple overlapping or non-overlapping continuous patches (each patch covering multiple time steps). Then, a linear embedding layer maps each patch to a compact patch embedding feature, transforming the original point-by-point sequence into a discrete patch token sequence, thereby significantly compressing the sequence length while preserving local contextual information. Then, the patches are embedded into the feature input residual quantum token hybrid module, which adopts a two-stream architecture of "parallel classical linear anchors and quantum nonlinear corrections, residual fusion". The classical linear anchors are used to capture the macroscopic linear trends in the grid load that are easy to fit and provide a stable performance lower bound for the model, while the quantum nonlinear correction path accurately extracts the complex nonlinear fluctuations and chaotic dynamics between patches caused by extreme weather or sudden power consumption through the high-dimensional mapping capability of the variable quantum circuit. After the two output residuals are added together, the gradient vanishing problem that is prone to occur in long sequence training of pure deep networks is avoided. At the same time, the strong nonlinear expression capability of quantum computing is used to enhance the accuracy of characterizing long-range dependencies across time periods. Finally, an enhanced feature sequence with richer semantics and stronger temporal representation capability is output, which lays a solid foundation for subsequent accurate mapping of future predictions of each sub-band.

[0050] As one example, the Residual Quantum Token Mixer (RQTM) is as follows: To address the issues of highly nonlinear cross-patch dependencies and the tendency of classical deep models to fall into the fitting-generalization dilemma in patch-based long sequence modeling, this invention designs RQTM. This module employs a two-stream residual architecture of "classical linear anchors and quantum nonlinear corrections" to explicitly perform frequency division modeling of inter-patch sequences. While ensuring training stability and robustness, it introduces the high-order expressive power of quantum models for complex dynamic modes.

[0051] The residual quantum token hybrid module employs a dual-stream architecture that combines parallel processing of classical linear mappings and quantum nonlinear mappings with residual fusion. The dual-stream architecture is as follows: Given input Patch representation ,in, This indicates the merged batch, feature, and channel dimensions. The input is the length of the patch sequence. RQTM models the sequence using parallel classical linear and quantum nonlinear flows, and its final output is defined as:

[0052] in, This represents the output of a classic linear anchor point. This is the correction term for the nonlinear residual in the quantum branch.

[0053] 1) Classical Linear Anchor: Considering the significant linear trends or low-frequency components often present in time series, the classical linear mapping is retained as the residual anchor in RQTM:

[0054] in, , This is used to predict the length of the patch sequence. This linear anchor provides a clear lower bound on the model's performance, allowing it to degenerate into a robust linear predictor even when quantum branch training is insufficient or gradient instability occurs. Simultaneously, this path provides an efficient gradient channel for backpropagation, helping to alleviate the optimization difficulties of quantum circuits in the early training stages.

[0055] 2) Quantum Non-linear Correction Path: The quantum branch aims to model high-order nonlinear residuals that are difficult to capture by classical linear layers. Its processing flow consists of three stages: feature compression, variable quantum evolution, and feature reconstruction.

[0056] (1) Feature compression and angle encoding: In order to transform the classical patch sequence into a quantum state that can be processed by variable quantum circuits, this invention first maps the input features into a latent representation that adapts to the parameters of the quantum rotation gate through a learnable nonlinear projection layer:

[0057] in, This step maps the cross-patch sequence to angular parameters that satisfy the periodic constraints of the quantum rotation gate.

[0058] Then, adopt The rotating door encodes the compressed features at an angle to obtain the initial quantum state:

[0059] (2) Variable quantum evolution: Based on the coding state, stacking Layer parameterization of strongly entangled layers to evolve quantum states:

[0060] in, These are trainable quantum parameters. This structure enhances the ability to represent complex periodic and chaotic fluctuation patterns by introducing multi-qubit entanglement, allowing the low-dimensional input to expand in an exponential-dimensional Hilbert space.

[0061] (3) Quantum measurement and feature reconstruction: After the evolution is completed, each qubit is measured under the Pauli-Z matrix (Pauli-Z) basis, and the expected value is extracted as the quantum feature:

[0062] This forms quantum eigenvectors. .

[0063] Eigenvectors obtained after quantum measurement The evolution of the input sequence in the quantum feature space is characterized.

[0064] To be fused with the output of the classic linear anchor in the same prediction space, the feature is projected onto the target patch prediction dimension via a classic linear mapping.

[0065] Specifically, the output of the quantum branch is represented as:

[0066] in, , .

[0067] This linear mapping structurally aligns the representations between the quantum feature space and the classical prediction space, enabling the output of the quantum branch to be combined with the classical linear anchor in residual form.

[0068] In S4, the time-series representation enhanced by the residual quantum token hybridization module has fully extracted and encoded cross-time-period dependencies in the high-dimensional latent space. However, the features at this point are still in an abstract high-dimensional embedding space, requiring decoding into specific future prediction values ​​corresponding to each frequency sub-band, and ultimately fusion into a transformer state prediction result with physical meaning. Traditional methods often directly use a simple fully connected layer to map features to prediction values ​​in the final prediction stage, but neglect the orthogonality and consistency constraints that each sub-band must satisfy during reconstruction after wavelet decomposition, leading to deviations in the physical dimensions of the prediction results or mismatches after fusion of prediction values ​​from different frequency bands. To this end, this invention first inputs the enhanced time-series representation into a linear prediction head, linearly mapping it from the high-dimensional embedding space to a sequence of future predicted values ​​corresponding to each frequency sub-band. Then, it performs an inverse discrete wavelet transform on the predicted values ​​of all sub-bands, utilizing the orthogonality and completeness of the wavelet transform to accurately reconstruct the prediction results of each frequency band into a complete multi-scale fusion sequence. Finally, it restores the original statistical distribution and physical dimensions of the data through inverse reversible instance normalization, outputting the transformer's future oil temperature and load prediction results with practical physical meaning. This reconstruction mechanism ensures the consistency and physical completeness of the prediction results of each frequency band during the fusion process, avoiding accumulated errors caused by frequency band mismatches. The power grid dispatch center can use this to identify the risk of oil temperature exceeding the threshold in advance, activate standby transformers, or perform load transfer, achieving a closed loop from data acquisition to forward-looking dispatch.

[0069] Specifically, in Q-WaveMixer, the time-series representation enhanced by the Residual Quantum Token Mixer (RQTM) has already fully encoded cross-time-period dependencies in a high-dimensional latent space. This feature is first input into the Linear Prediction Head, a learnable linear mapping layer, which projects it from the high-dimensional embedding space to the prediction dimension corresponding to each frequency sub-band. It outputs the point-by-point prediction value of each sub-band within the future prediction window (e.g., the next 24 hours, corresponding to 96 time steps), thereby decoding the abstract semantic features into frequency band prediction sequences with specific numerical meanings.

[0070] Then, the prediction results of all sub-bands enter the reconstruction stage: first, the inverse discrete wavelet transform (IDWT) is performed, and the prediction values ​​of each frequency band are synchronously fused by utilizing the orthogonality and completeness of the wavelet basis to reconstruct a complete multi-scale time domain sequence.

[0071] Next, inverse reversible instance normalization (inverse RevIN) is performed to restore the reconstructed sequence from the normalized statistical distribution to the original physical dimensions and distribution range. Finally, a transformer future oil temperature and load prediction sequence with actual physical meaning is output, which the power grid dispatch center uses for early warning and load dispatching decisions.

[0072] In addition, to effectively train the hybrid architecture and ensure that the quantum module learns features that conform to physical laws, this invention designs a Physics-Aware Quantum Hybrid Loss function. It includes task loss and three physical constraints.

[0073] Specifically, the physical perception integrated loss function, based on the standard prediction error loss, jointly introduces channel correlation consistency constraints (forcing the topological structure between variables learned by QCE to align with the real physical coupling relationship), sparsity constraints (eliminating redundant channels), and time series smoothing constraints (suppressing quantum noise), thereby guiding the parameters of the hybrid quantum-classical neural network to converge to the optimization region that conforms to physical laws, effectively improving the training stability and generalization performance of the model in noisy environments.

[0074] Robust Task Loss: Utilizes Huber Loss as the primary task loss. To enhance robustness to outliers in time series data:

[0075] Quantum Channel Correlation Consistency: To ensure that QCE correctly reflects the physical dependencies among multiple variables, this invention mandates quantum attention weights. autocorrelation matrix With real data cross-correlation matrix Maintain consistency:

[0076] in, Reflects quantum weight and The similarity between the physical data and the topology is required. This constraint forces the VQC to learn a topology that matches the real topology of the physical data.

[0077] Sparsity and Smoothness Constraints: This invention introduces... Apply QCE weights Regularization is used to remove redundant channels; and regularization is introduced. A total variation constraint is applied to the quantum output of the RQTM to suppress high-frequency quantum noise:

[0078] This multi-objective optimization strategy effectively solves the overfitting problem of Hybrid Quantum-Classical Architecture (HQCA) and improves the interpretability of the model.

[0079] This invention also included experimental design. To comprehensively evaluate the model's generalization ability and robustness across different domains, extensive experiments were conducted on six mainstream multivariate time series datasets covering fields such as energy management and weather forecasting. These datasets have different sampling frequencies, variable dimensions, and data sizes, which can fully verify the model's performance in handling long-term dependencies, seasonal fluctuations, and data heterogeneity. Detailed statistical characteristics are shown in Table 1.

[0080] Table 1. Dataset Statistics

[0081] As shown in Table 1, the datasets include the number of variables (C), the number of samples in each dataset partition, and the sampling frequency; the specific characteristics of each dataset are as follows: 1) ETT (Electricity Transformer Temperature): This dataset, introduced by Zhou et al., records the oil temperature and load characteristics of power transformers in two different regions of China, spanning from July 2016 to July 2018. Based on the sampling frequency, the dataset is subdivided into four subsets: hourly datasets (ETTh1, ETTh2) and 15-minute datasets (ETTm1, ETTm2).

[0082] 2) Weather: This dataset, processed and published by Wu et al., contains 21 meteorological indicators (such as temperature, humidity, and wind speed) recorded by the Max Planck Institute for Biogeochemistry in Germany. The data was sampled at 10-minute intervals, covering climate change records for the entire year of 2020.

[0083] 3) Electricity: This dataset originates from the UCI Machine Learning Library and was compiled into the LTSF benchmark by Wu et al. It contains hourly electricity consumption records (in kWh) for 321 clients, spanning from 2012 to 2014, and is a typical high-dimensional multivariate time series dataset.

[0084] When setting up the experiment, the standard LTSF (Long-term Time Series Forecasting) experimental protocol was followed, and all datasets were strictly divided into training set, validation set and test set in chronological order.

[0085] 1) Data partitioning: The ETT dataset is partitioned in a 6:2:2 ratio, while the Electricity and Weather datasets are partitioned in a 7:1:2 ratio.

[0086] 2) Prediction window: Following the standard protocol in the LTSF field, the prediction step size is set to... To cover forecasting needs across different time spans.

[0087] This invention uses mean squared error (MSE) and mean absolute error (MAE) as core evaluation indicators. Lower values ​​indicate higher prediction accuracy. The calculation formulas are as follows:

[0088] in, They represent The true value vector and the predicted value vector at each moment, To predict the window length, and They represent Norm and Norm.

[0089] To comprehensively verify the superior performance of Q-WaveMixer, this invention compares it with 11 representative state-of-the-art (SOTA) models. These models cover classic linear models, mainstream Transformer variants, and the latest MLP / CNN / Mixer hybrid architectures: MLP / CNN / Mixer architecture: WPMixer (AAAI 2025): This architecture combines wavelet transform and patching, and is an important comparative object of the method in the frequency domain mixing direction of this invention.

[0090] TimeFilter (ICLR 2025): This paper focuses on the latest research on frequency domain filtering and time feature decoupling, demonstrating its advantages on complex periodic data.

[0091] CCM (NeurIPS 2024): Proposed an innovative channel clustering strategy that groups channels by mining the similarity between variables, thereby reducing redundancy and enhancing multivariate dependency modeling.

[0092] ModernTCN (ICLR 2024): It draws on the design concepts of modern ConvNet (such as large kernel convolution) and achieves long sequence prediction performance comparable to Transformer while maintaining efficient computation.

[0093] TimeMixer (ICLR 2024): Employs a multi-scale hybrid architecture (Past-Decomposable-Mixing) to effectively utilize downsampling information to handle long-term dependencies.

[0094] MSGNet (AAAI 2024): A lightweight model based on multi-scale grid interaction that achieves competitive accuracy while maintaining high computational efficiency.

[0095] DLinear (AAAI 2023): A simple linear model incorporating trend decomposition, serving as an important baseline for testing whether complex deep models are overfitting.

[0096] Transformer-based architecture: MultiPatchformer (2025): The latest patch-based variant, which further optimizes the granularity of feature extraction through multi-scale patch interactions. It is abbreviated as MPformer.

[0097] iTransformer (ICLR 2024): Employs an inverted dimension strategy, applying an attention mechanism in the time dimension to capture multivariate correlations.

[0098] PatchTST (ICLR 2023): Introduced patching technology and channel independence strategy, which greatly improved the Transformer's ability to process long sequences.

[0099] Crossformer (ICLR 2023): Utilizes two-stage Attention (TSA) and cross-dimensional mechanisms to explicitly capture the dependencies between time series and variables.

[0100] The model implementation of this invention is based on the seamless integration of the PennyLane quantum library and the PyTorch deep learning framework, supporting end-to-end backpropagation of quantum-classical hybrid gradients. All experiments were performed on a single NVIDIA GeForce RTX 5060Ti (16GB) GPU. Given that existing work has limitations in input length (…), Differences in input length and hyperparameter configuration often make direct comparisons with original literature reports unfair. Therefore, to comprehensively evaluate the performance boundaries and generalization ability of Q-WaveMixer, this invention conducted comparative experiments under standardized settings. Specifically, in the standardized settings, the input length of all baseline models was fixed at [value missing]. This experiment was conducted in three independent replicates, and the average results were reported to eliminate randomness.

[0101] The experimental results and analysis are as follows: 1. Results and Analysis of the Main Comparative Experiment To comprehensively evaluate the performance of Q-WaveMixer, this invention conducted standardized comparison experiments with a uniform input length setting. The results are summarized in Table 2: Standardized Comparison: Multivariate Long-Term Prediction Results on 6 Mainstream Datasets. The input length of all models was uniformly fixed. L =96. Optimal results are bolded, and suboptimal results are underlined. Results for models marked (*) are taken from WPMixer, TimeFilter and MSGNet results are taken from TimeFilter, and others are taken from their respective models.

[0102] Standardization Comparison Analysis: As shown in Table 2, under standardization settings (all models have uniform input length), Under these conditions, Q-WaveMixer demonstrated superior overall predictive capabilities: Multivariate long-term prediction results on six mainstream datasets. All model input lengths are uniformly fixed. L =96. Optimal results are bolded, and suboptimal results are underlined. Results for models marked (*) are taken from WPMixer, TimeFilter and MSGNet results are taken from TimeFilter, and others are taken from their respective models.

[0103] Table 2: Standardized Comparison Experiment

[0104] Overall performance advantage: In 24 prediction settings covering 6 datasets, Q-WaveMixer ranked first or second in the vast majority of cases. Especially on the ETTh1, ETTh2, and ETTm2 datasets, Q-WaveMixer achieved the lowest MSE error in almost all prediction windows, and its overall performance is superior to the comparison models.

[0105] Comparison with state-of-the-art frequency domain filtering: On datasets with strong physical coupling between variables such as Weather and Electricity, although TimeFilter demonstrates strong competitiveness (outstanding in some metrics), Q-WaveMixer remains close behind, maintaining a sub-optimal performance. This indicates that the QCE module has comparable capabilities to state-of-the-art frequency domain filtering methods in capturing high-frequency nonlinear dependencies, and exhibits stronger generalization on the ETT dataset.

[0106] Long sequence robustness: Notably, as the prediction window extends to the most challenging... Q-WaveMixer demonstrates long-term stability. Experimental data shows that it performs well on all six datasets. In the step-size prediction setting, the model and its variants all achieved the best or second-best results. This result demonstrates that even over extremely long time spans, the architecture can effectively suppress error accumulation and maintain the accuracy of feature extraction, exhibiting stronger resistance to decay and generalization ability compared to the comparison models.

[0107] Ablation experiments were also conducted. To verify the effectiveness of each core component in Q-WaveMix, this invention used three representative datasets—ETTh1, ETTm1, and Weather—with all model input lengths uniformly fixed at [length value missing]. For prediction windows Five variant models were constructed for ablation experiments. Detailed experimental results are as follows: Figures 4-5 As shown. Figure 4 The bar chart compares the mean squared error (MSE) of ablation experiments on the ETTh1, ETTm1, and Weather datasets (prediction window T=96). The red bars represent the full Q-WaveMixer model, which has the lowest MSE on all three datasets (approximately 0.363, 0.309, and 0.159 respectively), demonstrating the best performance. The MSE of each variant model (cool-toned bars) is higher than that of the full model. Removing the classic residual path (w / o RQTM-C, gray bars) results in the largest increase in error, validating its crucial role in training stability. Removing the QCE module (blue bars) shows significant degradation on the physically tightly coupled Weather dataset. Removing wavelet decomposition (purple bars) significantly increases error on the highly non-stationary ETTh1 dataset. Removing QHL loss (green bars) shows consistent performance degradation across all datasets. Figure 5A bar chart compares the MAE (Mean Absolute Error) of ablation experiments on the ETTh1, ETTm1, and Weather datasets (prediction window T=96). Consistent with the MSE trend, the full Q-WaveMixer model (red bars) maintains the lowest MAE on all datasets, at 0.392, 0.338, and 0.196 respectively. All variant models have higher MAE than the full model, with the most significant performance degradation occurring from removing the classic residual path (w / o RQTM-C, gray bars); removing QCE (blue bars) shows the largest increase in MAE on the Weather dataset; removing wavelet decomposition (purple bars) performs poorly on ETTh1; and removing QHL (green bars) shows a consistent increase in MAE across all datasets, further confirming the indispensability of each core component to the model's prediction accuracy.

[0108] Results analysis: w / o QCE: The quantum channel entangler is removed and replaced with the classic Squeeze-and-Excitation module. As shown by the blue bars in the figure, the model's error is higher than the full model (red bars) on all datasets. In particular, the performance drops significantly on the Weather dataset, where the variables are physically coupled, which intuitively shows that quantum entanglement provides additional gains in capturing complex physical couplings.

[0109] w / o RQTM-Q (Quantum Path): Removing the quantum nonlinear flow from RQTM degenerates the model into a purely linear network. Experiments show a slight but consistent decrease in prediction accuracy (light blue bars are slightly higher than red bars), indicating that the quantum nonlinear mapping helps capture residual details in sequences that linear models cannot fit.

[0110] w / o RQTM-C (Classical Anchor): Removes the classic residual path from RQTM. This variant (gray bars) shows the most significant error increase in the graph, which visually strongly validates the key role of the classic anchor in stabilizing hybrid architecture training and mitigating gradient vanishing.

[0111] w / o Wavelet: Removes the multi-scale wavelet decomposition at the input. On the ETTh1 dataset, which has strong non-stationarity, this variant (purple bars) shows a significant increase in prediction error, demonstrating the necessity of multi-scale frequency domain decoupling for modeling non-stationary sequences.

[0112] w / o QHL: Training was performed using only the standard MSE loss. The green bars show that removing this component resulted in a consistent performance degradation across all datasets. This confirms that QHL is not merely used to aid convergence, but plays a crucial regularization role: by imposing physical consistency constraints, it effectively reduces the invalid search space for quantum parameters, prevents VQC from overfitting to high-frequency noise, and thus improves the model's generalization ability.

[0113] This invention also delves into the internal working mechanism, parameter efficiency, and robustness under extreme environments of Q-WaveMixer through visualization analysis and stress testing.

[0114] As the dimensions of multivariate time series increase (e.g., the number of channels in the Traffic dataset), The computational cost of capturing dependencies between variables has become a key bottleneck restricting the scalability of models. Table 3 compares the parameter complexity and receptive field of view of different models when dealing with channel dependencies.

[0115] Table 3. Comparison of the complexity of channel dependency modeling for different architectures.

[0116]

[0117] As shown in Table 3, among which, C For the number of channels (variable dimension). k The kernel size is denoted by . Q-WaveMix is ​​the only architecture that achieves linear parameter complexity while maintaining the global receptive field.

[0118] Analysis Conclusion: The bottleneck of classic global methods: Methods represented by iTransformer and MLP-Mixer use fully connected or full attention mechanisms to mix channel information. The number of parameters is proportional to the square of the number of channels (…). For high-dimensional data (such as...) This can lead to a surge in the number of parameters, which not only increases the computational burden but also makes it very easy to cause overfitting.

[0119] Limitations of Local and Independent Methods: Although Convolutional Neural Networks (CNNs) reduce complexity to linear... However, it can only capture local neighborhood relationships and cannot model long-distance variable coupling; although CI strategies such as PatchTST have achieved... While achieving high-efficiency computation, this comes at the cost of completely abandoning channel dependence, resulting in the loss of physical information.

[0120] A breakthrough in quantum architecture: The QCE module in Q-WaveMix breaks the aforementioned deadlock. Utilizing the parameter-sharing characteristics of VQC, this invention only requires... The parameters (i.e., the parameters of the quantum rotating gate) High-dimensional quantum states can be constructed using a linear relationship between the number of qubits. More importantly, through quantum entanglement, non-local global correlations can be established between qubits. This allows Q-WaveMix to achieve linear complexity. The system implements a global receptive field, thus exhibiting good parameter efficiency and generalization ability when processing high-dimensional datasets.

[0121] Considering the future deployment potential on NISQ hardware, this invention introduces a strength of [value missing] during the inference phase. Depolarization noise, in the ETTh1 dataset ( Taking this as an example, we simulate a real quantum computing environment. For example... Figure 6 As shown, the experiment revealed two distinctly different dynamic behaviors: Figure 6 In the ETTh1 dataset ( Noise robustness analysis on ). The horizontal axis represents the depolarization noise probability. The vertical axis represents MSE. The solid blue line shows that Q-WaveMixer did not crash under strong noise, but instead smoothly degraded to a level close to the linear baseline (dashed gray line) by utilizing the performance lower bound provided by the classic residual path.

[0122] Stability within the reliable region ( In the low-noise region (green area in the figure), the MSE of Q-WaveMixer only slightly increases from the optimal 0.358 to 0.366, demonstrating extremely high anti-interference ability. This is due to the regularization of the parameter space by the QHL loss function, which makes the model parameters converge to a flat minimum region.

[0123] Graceful degradation under strong noise ( As the noise probability increases to (Strong noise), the error of the pure quantum variant (dashed line) diverges sharply (MSE). Q-WaveMixer (solid line) exhibits a significant “graceful degradation” characteristic: its prediction error rises slowly and eventually converges to a level comparable to the linear benchmark (DLinear, MSE=0.397) (0.396).

[0124] This result experimentally confirms the design value of the Classical Linear Anchor in RQTM: as a stability anchor, this path provides a clear performance lower bound when the quantum circuit fails due to noise, thus ensuring the usability of the model in extreme environments.

[0125] To verify whether the QCE module captures the true physical dependencies through quantum entanglement, rather than merely fitting data noise, this invention performed a visualization analysis on the Weather dataset. This invention extracted the quantum entanglement matrix (based on cross-correlation reconstruction of measured values) from the trained QCE module and compared it with the Pearson correlation coefficient matrix (Ground Truth) of the original meteorological data. The results are as follows: Figures 7-8 As shown, this demonstrates a visualization of physical correlations on the Weather dataset.

[0126] Specifically, Figure 7 The Pearson correlation coefficient heatmap (Ground Truth) for the original meteorological indicators on the Weather dataset shows the distribution of strong physical correlations between real variables (such as "wet-bulb temperature" and "dew point temperature", "atmospheric pressure" and "visibility"). Figure 8 The quantum entanglement weight matrix learned by the QCE module accurately represents the highlighted blocks at the positions of the corresponding strongly physically correlated variable pairs, and... Figure 7 The high consistency of the actual physical distribution and the sparser and purer heatmap indicate that the quantum circuit successfully reconstructed the key physical coupling relationships and effectively filtered out spurious correlations.

[0127] In other words, certain variable pairs (such as wet-bulb temperature and dew point temperature, atmospheric pressure and visibility) exhibit naturally strong physical correlations in the Weather dataset. The QCE heatmap accurately displays high-activation blocks at the corresponding coordinate locations, demonstrating that the quantum circuit successfully simulates these nonlinear physical couplings using entangled states. Notably, compared to some weak background noise present in the Ground Truth, the QCE heatmap appears sparser and cleaner. This indicates that through the sparsity constraint in the QHL loss function, the model learns to automatically filter out spurious correlations that are useless for prediction, retaining only the key dynamic framework.

[0128] This result, from an experimental perspective, shows that the representation learned by Q-WaveMixer has good consistency with the known correlation structure of meteorological variables, demonstrating a certain degree of physical interpretability.

[0129] To verify whether the hybrid architecture effectively overcomes the "barren plateau" problem in quantum neural network training, this invention selected three representative datasets: ETTh1, ETTm1, and Weather, and analyzed them in the prediction window. Under the given settings, the validation set loss convergence trajectory during training was recorded and plotted, such as... Figure 9-11 As shown. Figure 9 In the figure, the gray dashed line (without classic anchor points) stagnates at 0.415 in the suboptimal flat region at Epochs 6-8, while the red solid line (complete model) converges rapidly to a lower value of 0.363, verifying the key role of the classic residual path in overcoming the barren plateau. Figure 10 The graph shows a comparison of training convergence dynamics on the ETTm1 dataset. The gray dashed line (removing the classic anchor point) gets stuck in the suboptimal flat region early on and can no longer decrease, while the red solid line (complete model) continues to converge smoothly to the optimal level of about 0.309, indicating that the architecture can maintain stable optimization dynamics on data with different sampling frequencies. Figure 11 The graph shows the training convergence dynamics on the Weather dataset. The gray dashed line (without the classic anchor point) stagnates at a suboptimal level of about 0.23, while the red solid line (complete model) converges rapidly to 0.159, which is significantly better than the variant model's 0.195. This demonstrates the ability of the design of this invention to successfully reshape and optimize the terrain on high-dimensional, strongly coupled datasets and guide the parameters into deeper global minima.

[0130] Specifically, Figure 9-11 Comparison of model training convergence dynamics on the ETTh1, ETTm1, and Weather datasets. The gray dashed line shows the model getting stuck in the suboptimal Barren Plateau after removing the classic anchor points; the orange dotted line (if any) shows oscillating convergence without QHL constraints; the red solid line shows that the full Q-WaveMixer model achieves fast and low-error convergence.

[0131] Evidence of the Barren Plateaus: such as Figure 9-11 As shown by the gray dashed line, after removing the classic linear anchor (without RQTM-C), the model exhibited significant premature convergence in the early stages of training (approximately Epochs 6-8). The loss curve stalled at a level significantly higher than the optimal solution (e.g., around 0.415 on ETTh1, about 14% higher than the optimal solution), and could hardly decrease further in subsequent Epochs. This objectively confirms that in the absence of gradient channels, deep VQC does indeed get stuck in a flat region of the parameter space, leading to gradient vanishing and ineffective updates.

[0132] Landscape Reshaping: In comparison, the full version of Q-WaveMixer (solid red line in the figure) exhibits superior dynamic characteristics. Thanks to the gradient path provided by RQTM and the physical regularization of QHL, the model not only overcomes the performance bottleneck of the gray curve but also demonstrates lower variance (smoother performance). This is especially true on the Weather dataset. Figure 11 The red curve converges rapidly to 0.159, significantly better than the variant model's 0.195. This indicates that the design of this invention successfully reshapes the optimized terrain, guiding the parameter search to a deeper global minimum region, rather than remaining at shallow local minima.

[0133] In summary, this invention addresses the highly non-stationary and physically tightly coupled characteristics of multivariate time series data, such as oil temperature and load, in power grid transformer operation monitoring. By employing a QCE module with the linear parameter complexity of quantum entanglement, it accurately reconstructs the thermodynamic and electrical global dependencies between oil temperature, current, voltage, and load, solving the problem of traditional models failing to balance global perception and parameter efficiency in channel modeling. Simultaneously, utilizing the dual-current residual architecture of RQTM, it employs a classical path to ensure gradient stability and a quantum path to capture nonlinear fluctuations caused by extreme weather or sudden power outages, effectively overcoming the contradiction between fitting capability and training robustness in long-sequence prediction. Combined with the physical consistency constraint of the QHL loss function, it ensures stable convergence and suppresses overfitting even with noisy real-world data. Validated on multiple real power grid datasets, this invention delivers high-precision transformer future state predictions for various prediction durations, providing dispatch centers with reliable forward-looking decision-making support for identifying oil temperature exceeding threshold risks, implementing load transfers, and activating standby transformers. Furthermore, this invention not only validates the effectiveness of oil temperature and load prediction on the Power Transformer Dataset (ETT), which covers strongly physically coupled data, but also demonstrates excellent generalization ability on publicly available benchmarks such as Weather and Electricity. Therefore, it can be widely applied to multivariate long-term prediction scenarios involving complex nonlinear dynamics, such as smart energy management, weather forecasting, and traffic flow control.

[0134] Example 2 In a typical embodiment of the present invention, a transformer state prediction system based on a hybrid neural network is provided, comprising: Data acquisition module: used to acquire raw multivariate time series data of transformer operation; Multi-scale decomposition and quantum channel recalibration module: used to decompose the transformer into low-frequency sub-bands reflecting the long-term evolution trend of the transformer and multiple high-frequency sub-bands reflecting sudden disturbances through discrete wavelet transform; each sub-band is input into a hybrid quantum-classical neural network, and the quantum channel entanglement module in the hybrid quantum-classical neural network recalibrates the global physical coupling relationship between multiple variables; The residual quantum temporal hybridization module is used to slice and linearly embed the recalibrated sub-band features to obtain patch embedded features; the patch embedded features are input into the residual quantum token hybridization module in the hybrid quantum-classical neural network to extract cross-time-period dependencies and obtain the enhanced temporal representation. Prediction output module: used to map the enhanced time series representation to the future prediction values ​​of each sub-band through the linear prediction head, and obtain the transformer state prediction results after reconstruction.

[0135] Example 3 The purpose of this embodiment is to provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the steps of the method of Embodiment 1.

[0136] Example 4 The purpose of this embodiment is to provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method of Embodiment 1.

[0137] The steps and methods involved in Embodiments 3 and 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0138] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0139] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0140] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A transformer state prediction method based on hybrid neural networks, characterized in that, include: Obtain the raw multivariate time series data of transformer operation; The transformer is decomposed into a low-frequency sub-band reflecting the long-term evolution trend of the transformer and multiple high-frequency sub-bands reflecting sudden disturbances through discrete wavelet transform. Each sub-band is then input into a hybrid quantum-classical neural network, where the quantum channel entanglement module of the hybrid quantum-classical neural network recalibrates the global physical coupling relationship between multiple variables. The recalibrated sub-band features are sliced ​​and linearly embedded to obtain patch embedded features; The patch embedding features are input into the residual quantum token hybrid module in the hybrid quantum-classical neural network to extract cross-time-period dependencies and obtain the enhanced temporal representation. The enhanced time series representation is mapped to the future prediction values ​​of each sub-band through a linear prediction head, and the transformer state prediction result is obtained after reconstruction.

2. The transformer state prediction method based on a hybrid neural network as described in claim 1, characterized in that, The original multivariate time series data includes at least transformer oil temperature, three-phase current, three-phase voltage, and active load.

3. The transformer state prediction method based on a hybrid neural network as described in claim 1, characterized in that, The quantum channel entanglement module aggregates time-dimensional information using global average pooling. After being compressed to the qubit dimension by linear projection, the physical coupling relationship between the multivariable channels is modeled by the strong entanglement layer and the two-qubit entanglement gate in the parameterized quantum circuit; The original features are recalibrated using channel attention weights.

4. The transformer state prediction method based on a hybrid neural network as described in claim 1, characterized in that, The quantum channel entanglement module divides the recalibrated features into discrete patch sequences according to the time span. Each patch contains multiple consecutive time steps, which are then mapped to patch embedded features through a learnable linear embedding layer.

5. The transformer state prediction method based on a hybrid neural network as described in claim 1, characterized in that, The residual quantum token hybrid module adopts a dual-stream architecture that processes classical linear mapping and quantum nonlinear mapping in parallel and fuses residuals. Among them, classical linear mapping captures the macroscopic trend of the sequence and serves as a stability anchor, while quantum nonlinear mapping extracts higher-order nonlinear fluctuations between patches through variable quantum circuits.

6. The transformer state prediction method based on a hybrid neural network as described in claim 5, characterized in that, The quantum nonlinear mapping encodes the patch embedding features into quantum rotation gate parameters via a learnable projection layer; After angular embedding and multi-layer strong entanglement evolution, quantum features are extracted using Pauli-Z basis measurement, and then a nonlinear correction term is output through linear mapping. This term is then added to the output residual of the classical linear mapping and normalized.

7. The transformer state prediction method based on a hybrid neural network as described in claim 1, characterized in that, The reconstruction includes sequentially performing inverse discrete wavelet transform to fuse the predicted values ​​of each sub-band, and performing inverse reversible instance normalization to recover the original physical dimensions of the prediction results, thereby obtaining the transformer state prediction results.

8. A transformer state prediction system based on a hybrid neural network, characterized in that, include: Data acquisition module: used to acquire raw multivariate time series data of transformer operation; Multi-scale decomposition and quantum channel recalibration module: used to decompose the transformer into low-frequency sub-bands reflecting the long-term evolution trend of the transformer and multiple high-frequency sub-bands reflecting sudden disturbances through discrete wavelet transform; each sub-band is input into a hybrid quantum-classical neural network, and the quantum channel entanglement module in the hybrid quantum-classical neural network recalibrates the global physical coupling relationship between multiple variables; Residual quantum temporal mixing module: used to slice and linearly embed the recalibrated subband features to obtain patch embedded features; The patch embedding features are input into the residual quantum token hybrid module in the hybrid quantum-classical neural network to extract cross-time-period dependencies and obtain the enhanced temporal representation. Prediction output module: used to map the enhanced time series representation to the future prediction values ​​of each sub-band through the linear prediction head, and obtain the transformer state prediction results after reconstruction.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method as described in any one of claims 1-7.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method as described in any one of claims 1-7.

Images