Method for predicting atmospheric ozone concentration based on near-far transfer interaction of double quantum line

Abstract

The application discloses a method for forecasting atmospheric ozone concentration based on near and long-range transport interaction of double quantum circuit. First, the application uses the pollutant observation before the report and the numerical weather forecast to construct a dynamic directed transport graph. Then, the source city-transport starting time-feature tensor is divided into two categories of near field and long range, and is input into two parameterized quantum circuits through dimension reduction and amplitude coding. According to the quantum measurement result, the near field and long range transport messages are recovered, and are fused with the local chemical-meteorological state of the target city. The near and long range transport messages of the transport classical ozone prediction model and the ozone observation data are jointly distilled to train the two quantum circuits, and the future ozone concentration is predicted. The application can keep the accuracy of the transport classical ozone prediction model, and can express and process the high-dimensional transport interaction relationship of different spatial scales in a compact quantum manner, and can provide an implementation path for the calculation acceleration on the applicable quantum hardware.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary fields of atmospheric environment prediction, spatiotemporal artificial intelligence and quantum computing, and specifically relates to an atmospheric ozone concentration forecasting method based on near- and long-range transmission interaction via dual quantum circuits. Background Technology

[0002] Atmospheric ozone is influenced by photochemical reactions, temperature and humidity, boundary layer mixing, precursor emissions, and regional transport, exhibiting strong nonlinearity, strong temporal sequence, and cross-regional propagation characteristics. The near-surface concentration of atmospheric ozone in a single city is determined not only by local meteorological and chemical conditions but also by the transport of pollution from upwind cities at different wind speeds, directions, and altitudes. Therefore, neglecting inter-city connections during ozone pollution events caused by strong transport processes can lead to a systematic loss of important predictive information, resulting in biased ozone pollution peak forecasts. The applicant has also found through sensitivity experiments that incorporating physical transport characteristics into classical ozone forecasting models improves ozone forecasting errors, with the improvement increasing with transport intensity. Therefore, incorporating transport information into models is particularly important for forecasting urban atmospheric ozone pollution events dominated by strong transport processes.

[0003] However, in scenarios involving nationwide cities, multiple wind fields at various altitudes, multiple transport start times, multiple pollutants, multiple forecast lead times, or ensemble forecasts, incorporating transport information from other cities into classic forecast models requires explicitly generating and processing a high-dimensional interactive feature tensor representing "target city × source city × transport start time × characteristic channel." Furthermore, due to its interactivity, the computational cost typically increases approximately quadratically with the number of cities; even with sparse graphs, the computational cost still cumulatively increases with the number of neighboring cities, lag time, feature dimension, number of lead times, and number of ensemble members. Therefore, it is necessary to compress this high-dimensional interaction while maintaining transport forecast accuracy to accelerate forecasting.

[0004] To address the rapidly increasing computational complexity of ozone prediction methods, existing methods based on classical computing architectures typically rely on increasing the model parameter size or employing features such as dimensionality reduction and sparsification to reduce the computational burden. However, with the increasing number of cities, multi-scale transport processes, and multi-time-dependent prediction scenarios, it remains difficult to balance the ability to characterize the transport process with computational efficiency. In recent years, the development of quantum computing technology has provided a new computational paradigm for processing complex high-dimensional information. Quantum machine learning methods based on quantum superposition, quantum entanglement, and parameterized quantum circuits can achieve high-dimensional feature space representation through quantum state encoding and quantum feature mapping, providing a new technical approach for compressed modeling of complex nonlinear interactions. Since the inter-city transport relationships formed during urban ozone transport can be represented as high-dimensional correlation structures, they are well-suited to the ability of quantum circuits to express high-dimensional feature spaces. Therefore, how to construct a method that incorporates quantum computing modules into urban ozone transport prediction models, mapping near-field and long-range high-dimensional transport interaction information onto quantum circuits respectively, and providing a new method for rapid prediction of large-scale urban ozone pollution, is a problem that urgently needs to be solved. Summary of the Invention

[0005] This invention addresses the problems existing in the prior art by providing an atmospheric ozone concentration forecasting method based on near- and long-range transmission interaction using dual quantum circuits. By encoding and feature mapping transmission information at different scales through quantum circuits, the method effectively reduces explicit high-dimensional interaction computations while maintaining the information representation capability of the transmission process. This provides a new computational method for rapid prediction of ozone pollution in large-scale cities and offers a verifiable technical path for future quantum hardware acceleration.

[0006] To address the above technical problems, this invention provides the following technical solution: an atmospheric ozone concentration forecasting method based on near- and long-range transmission interaction via dual quantum circuits, comprising the following steps:

[0007] Step S1: Obtain pollutant concentration sequences for multiple cities prior to the reporting start time and numerical weather forecast sequences released at the reporting start time. Form the local chemical-meteorological status of each city by sharing a classic encoder, and form source city transmission messages for multiple transmission start times. Step S2: Based on the city distance, the direction from the source city to the target city, and the multi-height forecast wind field between the start time of transport and the target forecast time, construct a dynamic directed transport map of the source city, the start time of transport, and the target city; calculate the effective wind speed and the expected transport distance along the direction from the source city to the target city based on the dynamic directed transport map, and determine the physical transport edge weights by combining wind direction consistency, city distance attenuation, the arrival matching degree between the expected transport distance and the distance between cities, and the pollutant survival rate, and further refine them to obtain the dynamic edge weights; Step S3: Divide the source cities into near-field sets and far-field sets, and construct near-field high-dimensional transport tensors and far-field high-dimensional transport tensors respectively using the source cities to transport messages and dynamic edge weights; Step S4: Construct and train the classical ozone forecasting model, which aggregates the near-field high-dimensional transport tensor and the long-range high-dimensional transport tensor respectively, generates classical near-field transport messages and classical long-range transport messages, and forecasts the future ozone concentration of the target city; Step S5: After performing quantum encoding preprocessing on the near-field high-dimensional transport tensor and the long-range high-dimensional transport tensor, input them into the near-field quantum circuit and the long-range quantum circuit, respectively; Step S6: Measure the two quantum circuits and recover the quantum near-field transport message and the quantum long-range transport message respectively. Fuse the quantum near-field transport message and the quantum long-range transport message with the local chemical-meteorological state of the target city to predict the ozone concentration of the target city at the target time. Jointly train the dual quantum circuit transport model with the classical near-field transport message, the classical long-range transport message and the ozone observation value at the target time as supervision information. Step S7: After the dual quantum circuit transport model is trained, the quantum near-field transport message, the quantum long-range transport message, and the local chemical-meteorological state of the target city are used as model inputs to predict the future ozone concentration of the target city.

[0008] Furthermore, in step S2 above, the physical weights are determined by city distance attenuation, the consistency between the forecast wind vector and the direction from the source city to the target city, the effective wind speed along the direction from the source city to the target city, the arrival matching degree between the expected transport distance and the city distance, and the pollutant survival rate; the physical weights are used to represent the relative importance of different source city-transport start time-target city combinations to the ozone forecast of the target city.

[0009] Furthermore, the aforementioned dynamic edge weights are obtained by the following method: the physical edge weights are corrected using an edge feature correction network and normalized in the dimensions of source city and transport start time; the edge features include city distance, remaining transport time, wind speed along the wind, wind direction consistency, arrival ratio, arrival matching degree, pollutant survival rate, boundary layer height, and photochemical conditions.

[0010] Furthermore, in the aforementioned step S3, the near-field set and the long-range set are divided according to the city distance threshold; the near-field high-dimensional transport tensor and the long-range high-dimensional transport tensor respectively include the source city dimension, the transport start time dimension and the feature channel dimension, and their tensor elements are obtained by multiplying the potential transport message of the corresponding source city with the corresponding dynamic edge weight.

[0011] Furthermore, in step S4 above, training the classical ozone forecasting model includes:

[0012] S4.1 Summing the near-field high-dimensional transport tensor and the long-range high-dimensional transport tensor along the source city and transport start time dimensions respectively, to obtain the classic near-field transport message and the classic long-range transport message;

[0013] S4.2 Input the classic near-field transmission message, classic long-range transmission message and the local chemical-meteorological state of the target city into the classic ozone prediction module, and train the shared classic encoder, side feature correction network, source city message encoder and classic ozone prediction module according to the error between the classic ozone prediction value and the ozone observation value at the target time.

[0014] S4.3 After training is completed, the parameters of the classic ozone prediction model are fixed and transmitted.

[0015] Furthermore, in step S5 mentioned above, the quantum encoding preprocessing includes:

[0016] S5.1. Using Principal Component Analysis (PCA), the transmission messages of each source city are reduced in dimension and compressed to form quantum near-field transmission tensor and quantum long-range transmission tensor.

[0017] S5.2 Flatten the two quantum transport tensors and truncate or fill them with zeros to a value not exceeding 2. n For a vector of dimension, calculate the L2 norm of each flattened vector and normalize it using the L2 norm. Then, encode the two types of normalized vectors into two parameterized quantum circuits, each with n data qubits.

[0018] S5.3. The normalized L2 norm is retained as the overall strength information of the corresponding transport tensor.

[0019] Furthermore, in the aforementioned step S5.2, the data qubits of each parameterized quantum circuit include a source city register, a transmission start time register, and a feature register; a trainable single-qubit rotation gate is applied inside the register, and parameterized entanglement gates are applied between the source city register and the feature register, and between the transmission start time register and the feature register, to express the interaction relationship between the source city, the transmission start time, and the feature channel; the training parameters of the near-field quantum circuit and the long-range quantum circuit are independent of each other.

[0020] Furthermore, in step S6 above, the expected value of a single qubit of the near-field quantum circuit and the long-range quantum circuit, as well as the related expected value of the selected qubit pair, are measured respectively; the measurement results of each circuit, the L2 norm of the corresponding quantum input tensor before normalization, and the target city code are input into the classical message recovery layer to generate quantum near-field transport messages and quantum long-range transport messages with the same dimension as the classical near-field transport messages and the classical long-range transport messages, respectively.

[0021] Furthermore, the aforementioned near-field quantum circuit and long-range quantum circuit each employ 12 data qubits, of which 4 qubits constitute the source city register, 4 qubits constitute the transmission start time register, and 4 qubits constitute the feature register. The amplitude encoding of each circuit consists of a 4096-dimensional vector formed by 16 source city slots, 16 transmission start time slots, and 16 feature channels, outputting 82 quantum measurement results. The 82 quantum measurement results are concatenated with the 1D L2 norm and the 4D target city encoding to form an 87-dimensional vector, and then processed through a classical message recovery layer of 87→48→24 to generate a 24-dimensional quantum near-field transmission message or a 24-dimensional quantum long-range transmission message.

[0022] Furthermore, step S6 specifically includes: using a weighted sum of the errors between classical near-field transport messages and quantum near-field transport messages, the errors between classical long-range transport messages and quantum long-range transport messages, and the errors between the final ozone prediction value and the ozone observation value at the target time, to train and update the near-field parameterized quantum circuit, the long-range parameterized quantum circuit, and the corresponding classical message recovery layer; the data standardization parameters, dimensionality reduction parameters, edge feature standardization parameters, and message normalization parameters are fitted only by the training set.

[0023] Furthermore, step S7 specifically includes: inputting the near-field and long-range high-dimensional transport tensors constructed from the available multi-city historical pollutant observations, GFS forecast meteorological fields, and dynamic directed transport maps into the trained near-field and long-range quantum circuits, respectively. Through quantum state evolution, measurement, and message recovery, quantum near-field transport messages and quantum long-range transport messages are obtained. These are then fused with the local chemical-meteorological state of the target city and input into the already trained classical ozone prediction module to calculate the change in ozone concentration at the target forecast time relative to the initial forecast time. Based on the ozone concentration at the initial forecast time, the final forecast ozone concentration is obtained.

[0024] Compared with the prior art, the beneficial technical effects of the present invention using the above technical solution are as follows:

[0025] (1) The amplitude encoding of quantum circuits enables the state space of n data qubits to represent at most 2 n A normalized amplitude and structured entanglement are used to express the combined relationship between the source city, the transmission start time, and the mechanism features, thus providing a compact quantum representation for high-dimensional transmission interactions.

[0026] (2) The near-field quantum branch is used to characterize the near-field characteristics of the transport that are sensitive to the prediction accuracy, while the remote quantum branch is used to characterize the remote regional background information that can be approximated and is related to the transport. The two quantum lines together cover the entire transport process, which can greatly reduce the number of computational parameters and avoid the rapid increase in the model's computational load with the city size.

[0027] (3) Under the same input and training budget, the urban atmospheric ozone concentration forecast model using dual quantum circuits that considers the near-long-distance transport interaction process has a forecast accuracy comparable to that of the classical forecast model that considers the transport process, thus maintaining its accuracy; and it has better accuracy than the classical local forecast model that does not consider the transport process. Attached Figure Description

[0028] Figure 1 This is a general technical roadmap for ozone forecasting based on near- and long-range transmission interaction via dual quantum circuits;

[0029] Figure 2 It is a schematic diagram of physical prior edge weights, learnable correction model, and near-field and long-field transport partitioning;

[0030] Figure 3 It is a flowchart of input dimensionality reduction, amplitude encoding, parameterized quantum interaction, quantum measurement and message recovery for two quantum circuits: near-field quantum circuit and long-range quantum circuit.

[0031] Figure 4 This is a flowchart of message distillation and quantum-classical hybrid ozone forecasting, which transmits classical forecasting models and dual quantum circuit models.

[0032] Figure 5 This is a time-series comparison chart of ozone observation results with dual quantum circuits, classical forecast models for transport, classical forecast models for zero transport, and ozone observation results during the full sample and strong transport process. Detailed Implementation

[0033] To better understand the technical content of the present invention, specific embodiments are described below in conjunction with the accompanying drawings.

[0034] In this invention, various aspects of the invention are described with reference to the accompanying drawings, in which numerous illustrative embodiments are shown. Embodiments of the invention are not limited to those depicted in the drawings. It should be understood that the invention is implemented through any of the various concepts and embodiments described above, as well as the concepts and embodiments described in detail below, because the concepts and embodiments disclosed herein are not limited to any particular implementation. Furthermore, some aspects of the invention disclosed may be used alone or in any suitable combination with other aspects of the invention disclosed.

[0035] The following uses the 18th-hour ozone forecast for 13 cities in the Beijing-Tianjin-Hebei region as an example to illustrate the implementation process of this invention. The number of cities, distance threshold, feature dimension, and number of qubits in this embodiment can be adjusted according to the application area and computational conditions, and do not constitute a limitation on the scope of protection of this invention. For ease of understanding, in this embodiment, i represents the target city to be forecasted, j represents the source city that may affect the target city, t represents the start time of the forecast, H represents the forecast lead time, and L represents the historical length.

[0036] like Figure 1 As shown, the method of the present invention includes the following steps:

[0037] S1. Construct pollutant and weather forecast data samples with dual time axes. For the reporting start time t, historical length L, and forecast lead time H, obtain the multi-city pollutant concentration sequence C(t-L+1:t) before the reporting start time, and the global Forecast System (GFS) numerical weather prediction data sequence M(t:t+H) released at the reporting start time; and the ozone observation concentration O of the target city i at time (t+H). 3,i obs (t+H).

[0038] A shared classical encoder was used to extract historical pollutant concentration sequences and future weather forecast data sequences for each city. Photolysis, thermal, humidity, ventilation, boundary layer, and precursor mechanism characteristics were calculated to form the local chemical-meteorological state h for each city. i (t, H).

[0039] The study area includes Beijing, Tianjin, Baoding, Tangshan, Langfang, Shijiazhuang, Qinhuangdao, Zhangjiakou, Chengde, Cangzhou, Hengshui, Xingtai, and Handan. Historical pollutant data include O3, NO2, CO, SO2, and PM2.5. 2.5 and PM 10 The hourly urban concentrations are used; future weather data uses GFS forecasts already published at the time of reporting, including variables such as wind speed, wind direction, temperature, humidity, boundary layer height, solar radiation, and precipitation at different altitudes. Each sample has both historical and forecast timelines: historical pollutant data is only read up to the time of reporting t, while weather GFS data comes from numerical forecasts published at time t from time t to time t+H, with the ozone concentration at time t+H as the prediction target.

[0040] All variables are standardized. Historical pollutant concentration sequences are input into a shared chemical time-series encoder, and GFS forecast sequences are input into a shared future weather encoder. Sharing means that all cities use the same set of network parameters, thus avoiding the duplication of the same structure as the number of cities increases. Mechanistic features such as photolysis, thermal, humidity, ventilation, boundary layer, and precursors are added to the coding results. For example, solar radiation and sunshine duration are used to represent photolysis conditions, near-surface wind speed and boundary layer height are used to represent ventilation conditions, and NO2 and CO are used to represent precursor states. This information is combined with the city codes to obtain the local chemical-meteorological state h of target city i. i (t, H). Simultaneously, for different source cities and transmission start times, the historical chemical state, corresponding meteorological state, mechanistic characteristics, and city code of the source city are fused to generate potential transmission messages s for the source city. j,rWhere j represents the source city and r represents the start time of transmission.

[0041] S2. Construct a dynamic directed transport map. For target city i, source city j, and different transport start times r, calculate the distance between cities and the direction from the source city to the target city based on the city's latitude and longitude. Utilize the multi-altitude numerical weather forecast wind field between the transport start time and the target forecast time to calculate the effective wind speed and expected transport distance along the direction from the source city to the target city. Combine this with wind direction consistency, city distance attenuation, arrival matching degree between expected transport distance and city distance, and pollutant survival rate to determine the physical transport edge weights. Further refine the physical transport edge weights using boundary layer height, photochemical conditions, and remaining transport time, obtaining the dynamic edge weights of source city-transport start time-target city. The dynamic edge weights are used to represent the relative importance of each source city to the ozone forecast of the target city at different transport start times.

[0042] like Figure 2 As shown, in specific implementation, to represent the relative importance of different combinations of "source city-transport start time-target city" to the ozone forecast of the target city, the physical weights of any combination of target city i, source city j, and transport start time r are calculated. The calculation of physical weights first involves calculating the distance between cities and the direction from the source city to the target city based on the city's latitude and longitude. Then, the forecast wind fields at 10 m, 850 hPa, and 700 hPa are projected onto this direction to obtain the effective wind speed along the direction from the source city to the target city, and based on this, the distance the polluted air mass can travel in the remaining forecast time is estimated. The distance between cities is used to calculate distance attenuation, the wind field projection is used to determine whether the source city is located upwind of the target city, and the difference between the expected travel distance and the distance between cities is used to determine whether the polluted air mass can reach the target time. All of the above information is used to determine the physical weights. Source cities that are downstream, at a moderate distance, and can reach the target time have larger physical weights; those that are upstream or inaccessible have physical weights close to zero.

[0043] Based on the physical boundary weights, a learnable transport weight correction model is used for adjustment. This model utilizes nine boundary features, including city distance, remaining transport time, along-wind speed, wind direction consistency, arrival ratio, arrival matching degree, pollutant survival rate, boundary layer height, and photochemical conditions, to adjust the physical boundary weights. Its core relationship can be expressed as:

[0044] ,

[0045] Here, softmax is a normalization function. To fix the physical edge weights, To prevent small positive numbers from having an undefined logarithm, For a small correction network of 9→16→1, It has nine edge features. The dynamic edge weights are ultimately normalized.

[0046] S3. Divide near-field transport into near-field and long-field transport and construct a high-dimensional transport tensor. Based on a city distance threshold, divide the candidate source cities for each target city into near-field and long-field sets, with the preferred city distance threshold being 250 km. Construct near-field high-dimensional transport tensors based on the historical status of the source cities, the transport start time, and the chemical-meteorological characteristics of the source cities. and remote high-dimensional transport tensor .

[0047] refer to Figure 2 In this embodiment, the transport relationships are divided into near-field and long-range categories based on the distance between the source city and the target city. A threshold of 250 km is used: source cities within 250 km are included in the near-field set, and source cities exceeding 250 km are included in the long-range set. High-dimensional tensors of "source city × transport start time × characteristic channel" are generated for both near-field and long-range transport. and Each tensor element in both types of high-dimensional tensors is generated by the corresponding source city message s. j,r Multiply by dynamic edge weight We obtain the following. Both types of tensors use the same feature definition, are three-dimensional, and differ only in their spatial extent, representing transport from nearby cities and background transport from more distant areas, respectively.

[0048] S4. Based on the high-dimensional transport tensor, a classical ozone transport forecasting model is established. According to the dynamic directed transport graph, dynamic edge weights, and transport arrival lag time, information from different source cities, transport start times, and characteristic channels is interactively and weightedly aggregated to obtain classical near-field transport messages. and classic remote message delivery The two types of classic delivery messages are compared with the local chemical-meteorological conditions of the target city. i The (t, H) fusion is input into the classical prediction module to predict the future ozone concentration of the target city. This classical transport prediction model serves as the teacher model, and the classical near-field transport messages and classical long-field transport messages it generates serve as distillation supervision targets for subsequent dual-quantum circuits.

[0049] In this embodiment, inter-city transport information is considered in the classical ozone forecasting model, and the near-field high-dimensional tensor of target city i is... and remote high-dimensional tensors By summing along two dimensions—different source city j and transmission start time r—the near-field transmission messages for target city i can be obtained. and remote message transmission Subsequently, the local chemical-meteorological conditions of the target city will be analyzed. i (t, H) and near-field transport messages and remote message transmission Input the classic ozone prediction module and output the future change in ozone concentration. The final ozone prediction concentration output by the classic ozone forecast model is: .in, This refers to the ozone concentration in the target city at the time of reporting. To transmit the future ozone concentration predicted by classical forecasting models.

[0050] It will transmit future ozone concentrations predicted by classical forecasting models. Ozone observation value at the target time 3,i obs By comparing (t+H), the model is trained by minimizing the ozone prediction error, and the city state encoder, dynamic edge weight correction network, source city message encoder, and ozone prediction module are updated. After training, the parameters of the classic model are fixed, and its final generated near-field delivery messages are used. and remote message transmission As a distillation monitoring target for two quantum circuits.

[0051] S5. Perform quantum encoding preprocessing on the high-dimensional transport tensor. While maintaining consistency with the classical transport forecasting model in S4 regarding the target city, source city, transport start time, and input characteristics, extract the near-field three-dimensional transport tensor constructed in S3. and remote high-dimensional transmission of three-dimensional tensors This serves as the input for both near-field and long-range quantum circuits. The two tensors are subjected to dimensionality reduction, truncation or zero-padding, weighting, and normalization to transform each tensor into a form with a length not exceeding 2. n The real or complex vector is obtained and the resulting vector is loaded into the n data qubits of the corresponding quantum circuit through amplitude encoding.

[0052] like Figure 3 and Figure 4 As shown in the embodiment, in order to make the high-dimensional transport tensor suitable for quantum encoding, the near-field high-dimensional transport tensor is... and remote high-dimensional transport tensor The transmission source characteristic message s for each source city j at a single transmission start time r. j,r Dimensionality reduction was performed using principal component analysis (PCA) to reduce the dimensionality of s. j,r The dimensions are compressed to 16. Then, for each target city i in both the near-field and long-range quantum circuits, 16 source city slots, 16 transmission start time slots, and 16 feature channels are set, forming a quantum circuit input tensor with 16×16×16=4096 values. The L2 norm of this flattened input tensor is calculated. That is, the square root of the sum of the squares of all elements:

[0053] ,

[0054] Where, x q This is the q-th element after the tensor is flattened. Then, each element is divided by this L2 norm to obtain a normalized vector that satisfies the amplitude coding requirements. :

[0055] ,

[0056] Generally speaking, n qubits can amplitude-encode 2 n Therefore, the 4096-dimensional input in this embodiment requires 12 data qubits. The normalized 4096 values... These are used sequentially as the basis states for 12-qubit computation. The amplitude coefficients (q=0,…,4095) are used to prepare the quantum input state:

[0057] ,

[0058] Near-field and long-range information are encoded separately and input into the near-field quantum circuit and the long-range quantum circuit.

[0059] S6. Perform dual-quantum transport interaction. A trainable single-bit rotation gate is applied to the source city register, transport start time register, and feature register, respectively, by the near-field quantum circuit and the long-range quantum circuit. A parameterized entanglement gate is applied between at least two types of registers, which may include source city-feature, transport start time-feature, and source city-transport start time entanglement. The single-bit expectation value and selected multi-bit correlation expectation values ​​are measured by the dual quantum circuit. The measurement vectors generated by the quantum circuits are then used to recover the quantum near-field transport messages. and quantum teleportation Merge local chemical-meteorological conditions h i (t, H) is then input into the classical prediction module, which outputs the ozone concentration increment. The predicted ozone concentration for the target city at the target time is obtained. Two quantum circuits are jointly trained using classical near-field transmission messages, classical long-range transmission messages, and the ozone observation value at the target time as supervisory information.

[0060] In this embodiment, the near-field and long-range circuits each use 12 qubits, with identical structures but independent training parameters. Four qubits correspond to the source city slot, four to the transmission start time slot, and four to the feature channel. This allows the circuit structure to correspond to the three types of information in the input tensor.

[0061] Each line first performs a trainable single-qubit rotation, then establishes connections between different pieces of information through entanglement gates. The preferred structure includes 4 lines within the source city register, 4 lines within the transmission start time register, 6 lines within the feature register, 16 entangled edges between the source city and the feature, and 16 entangled edges between the transmission start time and the feature, for a total of 46 lines. Each branch has 82 quantum parameters, for a total of 164 parameters across two branches. The source city and the transmission start time interact indirectly through a shared feature register.

[0062] In this embodiment, the expected values ​​of 12 qubits in the X, Y, and Z directions are measured for each line, and the expected values ​​of the two-qubit correlations of 46 entangled edges are also measured, resulting in a total of 82 measurement results. These 82 measurement results, the 1D tensor flattening L2 norm, and the 4D target city code are concatenated to obtain an 87-dimensional vector. Then, through an 87→48→24 message recovery layer, the measurement results generated by the quantum lines are reconstructed into 24-dimensional quantum near-field transmission messages. and quantum teleportation The two types of quantum transmission messages are linked to the local state h of the target city. i Inputting (t, H) together into the classical prediction module, the output is the change in ozone concentration relative to the reporting time. The final result is the future ozone concentration of the target city:

[0063] .

[0064] in, This refers to the ozone concentration in the target city at the time of reporting. The future ozone concentration at time t+H is predicted by the dual-quantum circuit ozone prediction model.

[0065] A dual-quantum circuit is jointly trained using message distillation loss and ozone prediction loss. The constructed classical transport prediction model is used as the teacher model, and its generated classical near-field transport messages are processed. and classic remote message delivery The distillation supervision targets for the transport information output by the near-field quantum circuit and the long-range quantum circuit are respectively used. Simultaneously, the distillation errors of the near-field transport message, the long-range transport message, and the final ozone prediction error are minimized, allowing the dual quantum circuit to retain transport information effective for ozone prediction while compressing high-dimensional transport interactions. The training, validation, and test sets are split according to the target effective time, and the dual quantum circuit model is trained accordingly. All data standardization and dimensionality reduction processes are performed on the training set.

[0066] In this embodiment, a dual-quantum circuit is trained, and three objectives are simultaneously constrained during training: quantum near-field message delivery. It should be close to the classic near-field delivery message. Quantum remote message transmission It should approximate the classic remote message delivery method. The final ozone prediction by the dual-quantum circuit forecasting model It should approximate the actual observation. Therefore, the joint loss function used is as follows:

[0067] ,

[0068] In the formula, MSE is the root mean square error, and λ N , λ F , These represent the weighting coefficients for near-field transport message loss, far-field transport message loss, and ozone concentration prediction loss, respectively. In this embodiment, λ is taken as... N =1.5、λ F =1.0、 =0.25, which makes the model pay more attention to near-field transport information that has a greater impact on the prediction while learning the final ozone concentration.

[0069] The samples are sequentially divided into training, validation, and test sets according to the target effective time. Data standardization, PCA, edge feature standardization, and message normalization are all determined using only the training set data. The validation set data is used to select the number of training rounds and hyperparameters, and the test set data is used only for final evaluation. During training, the two quantum circuits and the message recovery layer after quantum circuit measurements are primarily updated.

[0070] S7. After the dual quantum circuit transport model is trained, input the near-field and long-field transport tensors corresponding to the time to be predicted to obtain the quantum near-field transport message and the quantum long-field transport message respectively. After fusing the two with the local chemical-meteorological state of the target city, input them into the trained ozone prediction module to output the ozone concentration change of the target city at the time to be predicted and obtain the ozone concentration.

[0071] In this embodiment, the trained dual-quantum-circuit transport model includes a near-field quantum circuit, a long-range quantum circuit, and a message recovery layer with pre-determined parameters. For new forecast start times and lead times, near-field and long-range high-dimensional transport tensors, constructed using available multi-city historical pollutant observations, GFS forecast meteorological fields, and dynamic directed transport graphs, are input into the trained near-field and long-range quantum circuits, respectively. Through quantum state evolution, measurement, and message recovery, quantum near-field transport messages and quantum long-range transport messages are obtained. These are then fused with the local chemical-meteorological state of the target city. The local chemical-meteorological state describes the target city's own baseline pollutant concentrations, photochemical reactions, thermal conditions, humidity, ventilation, and boundary layer conditions. The quantum near-field and long-range transport messages supplement the near- and long-range transport contributions from other cities. The fused state is input into the trained classical ozone prediction module to calculate the change in ozone concentration from the target forecast time to the start time. Based on the ozone concentration value at the start time, the final forecast ozone concentration is obtained. Thus, the trained dual-quantum circuit transport model is responsible for extracting and compressing cross-city near- and long-distance transport information, while the local chemical-meteorological state of the target city is responsible for characterizing local generation and removal conditions. Together, they complete the forecast of future ozone concentration in the target city.

[0072] Figure 5 This is a time-series comparison of the ozone prediction results from the dual-quantum circuit, the classical ozone prediction model considering transport, the classical ozone prediction model considering zero transport, and ozone observations during the full sample and strong transport processes. The graph shows that the MAE of the dual-quantum circuit for the full sample is 16.490, while the classical ozone prediction model considering transport processes has an MAE of 16.429, a difference of 0.061 (approximately 0.37%). In the sample with the highest transport intensity (10%), the MAE of the dual-quantum circuit is 18.499, while the MAE of the classical ozone prediction model considering transport processes is 18.550, indicating that the dual-quantum circuit performs slightly better. Therefore, the average error of the dual-quantum circuit in the full test sample and strong transport processes is close to or slightly better than that of the classical ozone prediction model considering transport processes, indicating that the dual-quantum circuit can still retain effective information for predicting strong transport processes even after significantly compressing the transport mapping. As a comparison of transport necessity, the classical zero-transportation forecast model, which does not consider inter-city transport processes, had MAEs of 17.817 and 20.122 in the full sample and strong transport sample, respectively, significantly worse than the dual-quantum circuit forecast model. Therefore, the dual-quantum circuit model can better track the temporal changes in ozone concentration. During periods of strong transport, its predictions are closer to the observed values ​​than the two classical models, demonstrating a better characterization ability of the ozone transport process.

[0073] Preferably, in the embodiment, the mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²) can also be reported for different numbers of cities, source cities, transmission start times, feature numbers, forecast lead times, and ensemble members. 2The model's prediction accuracy, parameter compression capability, and quantum acceleration potential are reflected in its strong transmission event error, parameter quantity, number of qubits, and training time.

[0074] While the present invention has been described above with reference to preferred embodiments, it is not intended to limit the invention. Those skilled in the art can make various modifications and refinements without departing from the spirit and scope of the invention. Therefore, the scope of protection of the present invention shall be determined by the claims.

Claims

1. A method for forecasting atmospheric ozone concentration based on near- and long-range transmission interaction via dual quantum circuits, characterized in that, Includes the following steps: Step S1: Obtain pollutant concentration sequences for multiple cities prior to the reporting start time and numerical weather forecast sequences released at the reporting start time. Form the local chemical-meteorological status of each city by sharing a classic encoder, and form source city transmission messages for multiple transmission start times. Step S2: Based on the city distance, the direction from the source city to the target city, and the multi-height forecast wind field between the start time of transport and the target forecast time, construct a dynamic directed transport map of the source city, the start time of transport, and the target city; calculate the effective wind speed and the expected transport distance along the direction from the source city to the target city based on the dynamic directed transport map, and determine the physical transport edge weights by combining wind direction consistency, city distance attenuation, the arrival matching degree between the expected transport distance and the distance between cities, and the pollutant survival rate, and further refine them to obtain the dynamic edge weights; Step S3: Divide the source cities into near-field sets and far-field sets, and construct near-field high-dimensional transport tensors and far-field high-dimensional transport tensors respectively using the source cities to transport messages and dynamic edge weights; Step S4: Construct and train the classical ozone forecasting model, which aggregates the near-field high-dimensional transport tensor and the long-range high-dimensional transport tensor respectively, generates classical near-field transport messages and classical long-range transport messages, and forecasts the future ozone concentration of the target city; Step S5: After performing quantum encoding preprocessing on the near-field high-dimensional transport tensor and the long-range high-dimensional transport tensor, input them into the near-field quantum circuit and the long-range quantum circuit, respectively; Step S6: Measure the two quantum circuits and recover the quantum near-field transport message and the quantum long-range transport message respectively. Fuse the quantum near-field transport message and the quantum long-range transport message with the local chemical-meteorological state of the target city to predict the ozone concentration of the target city at the target time. Jointly train the dual quantum circuit transport model with the classical near-field transport message, the classical long-range transport message and the ozone observation value at the target time as supervision information. Step S7: After the dual quantum circuit transport model is trained, the quantum near-field transport message, the quantum long-range transport message, and the local chemical-meteorological state of the target city are used as model inputs to predict the future ozone concentration of the target city.

2. The atmospheric ozone concentration forecasting method based on near- and long-range transmission interaction using dual quantum circuits according to claim 1, characterized in that, In step S2, the physical weights are determined by city distance attenuation, the consistency between the forecast wind vector and the direction from the source city to the target city, the effective wind speed along the direction from the source city to the target city, the arrival matching degree between the expected transport distance and the city distance, and the pollutant survival rate; the physical weights are used to represent the relative importance of different source city-transport start time-target city combinations to the ozone forecast of the target city.

3. The atmospheric ozone concentration forecasting method based on near- and long-range transmission interaction using dual quantum circuits according to claim 2, characterized in that, The dynamic edge weights are obtained by the following method: the physical edge weights are corrected using an edge feature correction network and normalized in the dimensions of source city and transport start time; the edge features include city distance, remaining transport time, wind speed along the wind, wind direction consistency, arrival ratio, arrival matching degree, pollutant survival rate, boundary layer height and photochemical conditions.

4. The atmospheric ozone concentration forecasting method based on near- and long-range transmission interaction using dual quantum circuits according to claim 1, characterized in that, In step S3, the near-field set and the long-range set are divided according to the city distance threshold; the near-field high-dimensional transport tensor and the long-range high-dimensional transport tensor include the source city dimension, the transport start time dimension and the feature channel dimension, respectively, and their tensor elements are obtained by multiplying the potential transport message of the corresponding source city with the corresponding dynamic edge weight.

5. The atmospheric ozone concentration forecasting method based on near- and long-range transmission interaction using dual quantum circuits according to claim 1, characterized in that, In step S4, training the classical ozone prediction model includes: S4.1 Summing the near-field high-dimensional transport tensor and the long-range high-dimensional transport tensor along the source city and transport start time dimensions respectively, to obtain the classic near-field transport message and the classic long-range transport message; S4.2 Input the classic near-field transmission message, classic long-range transmission message and the local chemical-meteorological state of the target city into the classic ozone prediction module, and train the shared classic encoder, side feature correction network, source city message encoder and classic ozone prediction module according to the error between the classic ozone prediction value and the ozone observation value at the target time. S4.3 After training is completed, the parameters of the classic ozone prediction model are fixed and transmitted.

6. The atmospheric ozone concentration forecasting method based on near- and long-range transmission interaction using dual quantum circuits according to claim 1, characterized in that, In step S5, the quantum encoding preprocessing includes: S5.

1. Using Principal Component Analysis (PCA), the transmission messages of each source city are reduced in dimension and compressed to form quantum near-field transmission tensor and quantum long-range transmission tensor. S5.2 Flatten the two quantum transport tensors and truncate or fill them with zeros to a value not exceeding 2. n For a vector of dimension, calculate the L2 norm of each flattened vector and normalize it using the L2 norm. Then, encode the two types of normalized vectors into two parameterized quantum circuits, each with n data qubits. S5.

3. The normalized L2 norm is retained as the overall strength information of the corresponding transport tensor.

7. The atmospheric ozone concentration forecasting method based on near- and long-range transmission interaction using dual quantum circuits according to claim 6, characterized in that, In step S5.2, the data qubits of each parameterized quantum circuit include a source city register, a transmission start time register, and a feature register; a trainable single-qubit rotation gate is applied inside the register, and a parameterized entanglement gate is applied between the source city register and the feature register, and between the transmission start time register and the feature register, to represent the interaction relationship between the source city, the transmission start time, and the feature channel; the training parameters of the near-field quantum circuit and the long-range quantum circuit are independent of each other.

8. The atmospheric ozone concentration forecasting method based on near- and long-range transmission interaction using dual quantum circuits according to claim 7, characterized in that, In step S6, the expected value of a single qubit in the near-field quantum circuit and the long-range quantum circuit, as well as the related expected value of the selected qubit pair, are measured respectively. The measurement results of each circuit, the L2 norm of the corresponding quantum input tensor before normalization, and the target city code are input into the classical message recovery layer to generate quantum near-field transport messages and quantum long-range transport messages with the same dimension as the classical near-field transport messages and classical long-range transport messages, respectively.

9. The atmospheric ozone concentration forecasting method based on near- and long-range transmission interaction using dual quantum circuits according to claim 8, characterized in that, The near-field quantum circuit and the long-range quantum circuit each use 12 data qubits, of which 4 qubits constitute the source city register, 4 qubits constitute the transmission start time register, and 4 qubits constitute the feature register. The amplitude encoding of each circuit is a 4096-dimensional vector formed by 16 source city slots, 16 transmission start time slots, and 16 feature channels, outputting 82 quantum measurement results. The 82 quantum measurement results are concatenated with the 1D L2 norm and the 4D target city encoding to form an 87-dimensional vector, and then generated into a 24-dimensional quantum near-field transmission message or a 24-dimensional quantum long-range transmission message through the classical message recovery layer of 87→48→24.

10. The atmospheric ozone concentration forecasting method based on near- and long-range transmission interaction using dual quantum circuits according to claim 1, characterized in that, Step S6 specifically includes: using the weighted sum of the errors between classical near-field transport messages and quantum near-field transport messages, the errors between classical long-range transport messages and quantum long-range transport messages, and the errors between the final ozone prediction value and the ozone observation value at the target time, to train and update the near-field parameterized quantum circuit, the long-range parameterized quantum circuit, and the corresponding classical message recovery layer; the data standardization parameters, dimensionality reduction parameters, edge feature standardization parameters, and message normalization parameters are fitted only by the training set.

Images