A short-term wind speed prediction method and system based on deep learning and quantum neural network, medium

Abstract

This invention belongs to the field of wind speed forecasting technology, specifically relating to a short-term wind speed forecasting method, system, and medium based on deep learning and quantum neural networks. The method involves acquiring time-series data required for short-term wind speed forecasting and inputting it into a trained short-term wind speed forecasting model corresponding to the season to which the time-series data belongs, thereby obtaining the short-term wind speed forecast result. The short-term wind speed forecasting model includes a feature extraction and alignment module and a wind speed forecasting module. The feature extraction and alignment module uses a deep learning network to extract features from the data input to the model and then align those features to map the current tensor to the quantum neural network space. The wind speed forecasting module uses the quantum neural network to forecast wind speed based on the aligned features. This model improves prediction accuracy by introducing deep learning theory; and by utilizing quantum neural networks, it achieves exponential high-dimensional data processing and computational acceleration with very few parameters, reducing the overall model optimization and training difficulty.

Description

Technical Field

[0001] This invention belongs to the field of wind speed prediction technology, specifically relating to a short-term wind speed prediction method and system, and a medium based on deep learning and quantum neural networks. Background Technology

[0002] As a crucial component of my country's energy system, wind power generation is widely applied in high-precision industries such as construction, automotive, military, and electronics. However, the intermittent, random, and uncontrollable nature of wind energy can severely impact the performance of wind turbines and energy storage devices, increasing safety risks and causing significant economic losses. Wind speed forecasting is a key technology in the wind power generation field, crucial for grid dispatching, wind farm operation and maintenance, and electricity market transactions. It is necessary to predict and analyze wind energy fluctuations over a future period. Short-term wind speed forecasting, in particular, directly affects the real-time fluctuation control of wind power, becoming a core element in ensuring the safe and stable operation of the power grid and enabling timely equipment adjustments to avoid economic losses.

[0003] Currently, short-term wind speed forecasting methods are mainly divided into two categories. One category is numerical weather prediction based on physical models, which predicts wind speed changes by solving atmospheric dynamic equations. However, this method involves large computational loads, is sensitive to initial and boundary conditions, and has limited accuracy in complex terrain and local micro-meteorological environments. The other category is data-driven statistical learning methods, including time series analysis (such as the autoregressive moving average model ARIMA), shallow machine learning (such as support vector machines SVM), and deep learning models (such as long short-term memory networks LSTM and convolutional neural networks CNN).

[0004] For example, Chinese invention patent application CN117878913A, published on April 12, 2024, discloses a method and system for predicting microscopic wind speed of wind turbines based on an ARIMA time series model. This method uses ARIMA modeling based on the microscopic wind speed time series data of a single wind turbine, and then uses the established model to predict wind speed, supporting intelligent wind power control for wind turbines to adjust operation, control power, and schedule maintenance plans within a very short period (a few minutes to tens of minutes). However, ARIMA is essentially linear modeling, assuming that future values ​​are a linear combination of historical values ​​plus a random error term. Atmospheric motion is a complex nonlinear chaotic system, and wind speed changes are affected by the nonlinear coupling of multiple factors. ARIMA cannot capture this nonlinear relationship, leading to large prediction errors when wind speed fluctuates drastically or changes abruptly. Moreover, once the model parameters are determined, they usually remain unchanged throughout the prediction process, while the statistical characteristics of wind speed change with meteorological conditions. ARIMA itself cannot dynamically adjust its structure to adapt to these changes, resulting in a decline in model performance.

[0005] For example, Chinese invention patent application CN120996266A, published on November 21, 2025, discloses a short-term wind speed prediction method based on the BDWPT-KAN model. This method leverages the powerful nonlinear mapping capability of the KAN model to construct a KAN model for wind speed prediction, effectively modeling the nonlinear laws of wind speed evolution. Using the KAN model for wind speed prediction can improve the prediction accuracy and generalization ability of the model during periods of rapid wind rise and fall. However, KAN focuses on function approximation. When dealing with wind speed sequences with significant temporal dependencies, its ability to capture long-term historical dependencies is weaker than dedicated time-series models, resulting in lower prediction accuracy. Furthermore, KAN requires both feature extraction and prediction; to fit complex wind speed changes, it needs deeper network layers or more nodes, leading to a larger number of adopted data and increased training difficulty. Summary of the Invention

[0006] The purpose of this invention is to provide a short-term wind speed prediction method, system, and medium based on deep learning and quantum neural networks, in order to solve the problems of low prediction accuracy and high training difficulty caused by using existing models for short-term wind speed prediction.

[0007] To address the aforementioned technical problems, the first aspect of this invention provides a short-term wind speed prediction method based on deep learning and quantum neural networks. The method includes: acquiring time-series data required for short-term wind speed prediction and inputting it into a trained short-term wind speed prediction model corresponding to the season to which the time-series data belongs, thereby obtaining a short-term wind speed prediction result; the short-term wind speed prediction model includes a feature extraction and alignment module and a wind speed prediction module; the feature extraction and alignment module is used to extract features from the data input to the model using a deep learning network and then align those features so that the current tensor is mapped to the quantum neural network space; the wind speed prediction module is used to predict wind speed using the aligned features via the quantum neural network.

[0008] In one possible implementation, the part of the deep learning network used for feature extraction includes convolutional neural network layers and fusion network layers, with the fusion network layers used to flatten the multidimensional inputs and then fuse the features.

[0009] In one possible implementation, the part of the deep learning network used for feature alignment is a gated recurrent neural network layer.

[0010] In one possible implementation, the required time series data includes historical meteorological feature time series data.

[0011] In one possible implementation, meteorological features include wind speed, air pressure, atmospheric temperature, and oxygen concentration.

[0012] In one possible implementation, the quantum input layer in a quantum neural network employs an angle data encoding strategy to convert the classical input information into a quantum information carrier.

[0013] In one possible implementation, the quantum output layer in a quantum neural network employs a quantum amplitude sampling strategy to convert quantum information into classical information for wind speed prediction.

[0014] In one possible implementation, the required data is preprocessed data, which includes one or both of denoising and missing value compensation processing.

[0015] In one possible implementation, the loss function for training the short-term wind speed prediction model includes a loss function that measures the deviation between the predicted and actual values ​​and a loss function that measures the goodness of fit of the model.

[0016] To address the aforementioned technical problems, a second aspect of the present invention provides a short-term wind speed prediction system based on deep learning and quantum neural networks, comprising a processor for executing a computer program to implement the steps of the method in any possible implementation of the first aspect of the present invention.

[0017] To address the aforementioned technical problems, a third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method in any possible implementation of the first aspect of the present invention.

[0018] The beneficial effects of this invention are as follows: This invention no longer uses existing models in the prior art, but designs a new short-term wind speed prediction model. This model first uses a deep learning network for feature extraction and feature alignment, and then uses a quantum neural network to predict wind speed based on the aligned features.

[0019] This model incorporates deep learning theory for feature extraction and alignment. This means the model can automatically learn the nonlinear mapping relationship between multi-source meteorological features and wind speed, enabling subsequent predictions to be based on purer and more representative features. During training, it significantly enhances global search capabilities and the ability to escape local optima. Its main advantages can be summarized as stronger global applicability and more stable performance, improving prediction accuracy. Furthermore, quantum neural networks utilize quantum superposition and parallelism to process information in parallel within a high-dimensional Hilbert space. This allows for exponential acceleration of high-dimensional data processing and computation with very few parameters, demonstrating strong potential in capturing the complex spatiotemporal dynamics that govern wind speed fluctuations. It also exhibits stronger fitting capabilities for the complex coupling relationships between multiple factors in wind speed prediction.

[0020] Overall, this invention adopts a cascaded architecture of deep learning and quantum neural networks. Compared with the existing technology where a single model has to undertake the dual tasks of feature extraction and prediction, it not only reduces the difficulty of overall model optimization and training, but also improves the robustness and interpretability of the system through modular design. It forms a technical route of first optimizing the input and then making efficient predictions, which can improve prediction accuracy and efficiency. Attached Figure Description

[0021] Figure 1 This is the overall network architecture diagram of the short-term wind speed prediction model based on deep learning and quantum neural networks of the present invention;

[0022] Figure 2 This is a quantum neural network structure diagram in the implementation of the short-term wind speed prediction method based on deep learning and quantum neural networks of the present invention;

[0023] Figure 3 This is a flowchart of the training process for the short-term wind speed prediction method based on deep learning and quantum neural networks of the present invention.

[0024] Figure 4 This is a comparison chart of the prediction results of the short-term wind speed prediction system based on deep learning and quantum neural networks of the present invention;

[0025] Figure 5 This is a structural diagram of the short-term wind speed prediction system based on deep learning and quantum neural networks of the present invention. Detailed Implementation

[0026] The core of this invention is the use of a cascaded architecture combining deep learning and quantum neural networks. Deep learning is used for feature extraction and alignment, while the quantum neural network is used to predict wind speed based on the aligned features. This approach significantly enhances global search capabilities and the ability to escape local optima during training, thereby improving prediction accuracy and reducing the overall model optimization and training complexity. By combining the stable feature extraction capabilities of deep learning with the powerful computational reasoning capabilities of quantum neural networks, the accuracy of wind speed prediction is greatly improved, solving the technical problems of low accuracy and poor generalization in existing wind speed prediction algorithms when performing sequential operations.

[0027] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings.

[0028] An implementation method for short-term wind speed prediction based on deep learning and quantum neural networks:

[0029] A short-term wind speed prediction method based on deep learning and quantum neural networks is proposed. This method mainly utilizes a short-term wind speed prediction model built on artificial intelligence to predict short-term wind speeds. The specific process is as follows:

[0030] Step 1: Obtain the historical dataset, which must include at least historical wind speed data.

[0031] The historical dataset in this step includes data from multiple consecutive time points. Each time point's data includes meteorological characteristics other than wind speed. Considering that wind speed is not determined by a single factor, as an example of meteorological characteristics, atmospheric temperature, air pressure, and oxygen concentration are included. These three meteorological characteristics can characterize the atmospheric motion state from different dimensions, providing information on the physical driving forces of wind speed changes for short-term wind speed prediction models. Wind is airflow caused by pressure gradients, which often originate from uneven heating of the Earth's surface. Temperature differences are the most fundamental energy source for atmospheric motion, so atmospheric temperature can indirectly reflect the trend of wind speed changes; therefore, atmospheric temperature is chosen here. As mentioned earlier, wind is directly driven by pressure gradient forces, with air flowing from high-pressure areas to low-pressure areas. The greater the pressure gradient, the greater the wind speed. A sudden drop in air pressure often indicates approaching strong winds or severe weather; therefore, atmospheric pressure is chosen here. Oxygen is one of the main components of air, and changes in oxygen concentration are closely related to air density. Air density directly determines the magnitude of wind energy, and changes in density are often accompanied by adjustments in airflow characteristics; therefore, oxygen concentration is chosen here. Based on the above meteorological characteristics, we attempt to enable short-term wind speed prediction models to not only know the past wind speeds, but also the thermal and dynamic states of the atmosphere, thereby making predictions that are more in line with physical laws.

[0032] Due to various factors such as the observation environment (e.g., lightning interference), equipment limitations (e.g., equipment aging), and data transmission (e.g., line aging), the data obtained from continuous time points often suffers from varying degrees of data loss and anomalies. This disrupts the continuity of the time series and poses challenges to subsequent model training and prediction accuracy. Therefore, as a better processing method, preprocessing of these data is necessary. Specific preprocessing can include denoising and / or missing value compensation (or missing value imputation). As an example of denoising, moving average filtering can be used to filter out random errors, high-frequency interference, or sensor noise in the observed data, retaining the trend components reflecting the real physical process; alternatively, the Kalman filter algorithm can be used to recursively estimate the true state using observed values ​​and system state equations. As an example of missing value imputation, interpolation methods such as linear interpolation and cubic spline interpolation can be used; statistical imputation methods can also be used, such as filling with the mean of the entire sequence or neighboring time periods. Appropriate preprocessing methods can be selected for different situations.

[0033] Furthermore, considering that wind speed exhibits seasonality and a periodic variation pattern consistent with seasonal changes over time, the data from the aforementioned multiple consecutive time points are divided into seasonal categories. This means that the data from each time point also carries a seasonal label, which helps the short-term wind speed prediction model improve prediction accuracy based on seasonal characteristics and enables predictions for specific time periods based on different network parameters.

[0034] Of course, after the above processing, the data needs to be standardized to eliminate the influence of units and unify the feature scale. This is a key prerequisite to ensure effective model training and stable prediction.

[0035] Step 2: Build a short-term wind speed forecasting model.

[0036] Short-term wind speed forecasting models are the core of wind speed forecasting, and their overall framework is as follows: Figure 1 As shown, the short-term wind speed prediction model includes a feature extraction and alignment module and a wind speed prediction module. The feature extraction and alignment module uses a deep learning network to extract features from the data input to the model and then align the features so that the current tensor is mapped to the quantum neural network space. The wind speed prediction module uses a quantum neural network (based on quantum mechanics theory) to predict wind speed from the aligned features.

[0037] 1) Feature extraction and alignment module.

[0038] The feature extraction and alignment module is... Figure 1The "Deep Learning Module" in the model includes a feature extraction layer and a feature alignment layer. The feature extraction layer extracts features from the data input to the model to obtain a tensor containing rich feature information; the feature alignment layer performs feature alignment to map the current tensor to the quantum neural network space.

[0039] As one embodiment of the feature extraction layer, the feature extraction layer includes a convolutional neural network layer and an integration network layer.

[0040] Convolutional neural network layers can have several architectural schemes: The first scheme is a parallel architecture, in which the convolutional neural network layer includes multiple parallel convolutional branches. Each convolutional branch has a normalization layer, a convolutional layer, and a pooling layer connected in sequence. The normalization layer is used to normalize the input wind speed tensor. The convolutional layer is used to extract features from the normalized wind speed tensor to obtain a feature matrix. The pooling layer is used to spatially downsample the tensors in the feature matrix to reduce the spatial size of the feature map. The parameters of the convolutional layers on different convolutional branches are different to extract features of different scales and characteristics. The pooling layer can be an average pooling layer. Figure 1 The first approach is a parallel architecture, specifically consisting of three convolutional branches. The first branch contains a single-channel 4×4 convolutional layer, while the second and third branches each contain two 5-channel 4×4 convolutional layers with different strides. Except for the last convolutional layer, which uses ReLU activation, all other layers use Leaky ReLU activation. To maintain consistent feature map sizes and enhance spatial information preservation, zero-padding is applied in each convolutional operation. The second approach is a serial architecture, directly comprising a normalization layer, multiple sequentially connected convolutional layers, and pooling layers. The third approach is a hybrid architecture combining the serial and parallel approaches. The final architecture chosen depends on the specific circumstances.

[0041] Integration network layers are used to flatten multidimensional inputs for feature fusion, specifically including Flatten and Concatenate; for Figure 1 In the network, the outputs of the three convolutional branches are first flattened using Flatten and then concatenated using Concatenate.

[0042] As one embodiment of the feature alignment layer, a gated recurrent neural network (RNN) is chosen. This RNN layer aligns the extracted features to map the current tensor to the quantum neural network space. A gated RNN is a variant of a recurrent neural network (RNN), similar to a Long Short-Term Memory (LSTM) neural network. It addresses the limitations of long-term memory and gradient issues during backpropagation in RNNs through gating units. Compared to LSTM, the internal architecture of a gated RNN is simpler and computationally easier. Gated RNNs can filter out irrelevant noise and reduce overfitting through gating mechanisms, thus attenuating the noise impact of anomalous changes during propagation. This makes them particularly suitable for time series with clear seasonality and trends. A gated RNN contains two gates: an update gate and a reset gate. The reset gate determines how new input information is combined with previous memory, while the update gate defines the amount of previous memory stored at the current time step. Setting the reset gate to 1 and the update gate to 0 results in a standard RNN model. These two gating vectors determine which information is ultimately output by the gated recurrent unit. What makes these two gating mechanisms special is that they can preserve information in long-term sequences without it being erased over time or removed because it is no longer relevant to the prediction.

[0043] 2) Wind speed prediction module.

[0044] The wind speed prediction module is Figure 1 The "quantum neural network module" in the model uses a quantum neural network, which is the core prediction network of the wind speed prediction model. The construction of the quantum neural network is based on the two major concepts of quantum computing and artificial neural networks. It performs feature analysis and reasoning in the quantum neural network space based on the extracted relevant features to make wind speed predictions.

[0045] Quantum computing, with its unique computational characteristics, has become an ideal technical solution for improving the computational efficiency of neural networks. In quantum-inspired neural network models, quantum neural computing is usually based on the real number assumption, that is, all quantum probability amplitudes are real numbers. The vast majority of quantum-inspired neural network architectures are built on the qubit neuron model, and the transition between neuron states and quantum states is controlled by operations derived from quantum logic gates.

[0046] Figure 2The overall structure of the proposed quantum neural network is shown. Specifically, it includes a quantum input layer (Input(L)) for converting classical data into quantum information carriers, a quantum hidden layer (Hidden(M)) for performing quantum feature transformation and entanglement learning for feature analysis in quantum space, and a quantum output layer (Output(N)) for converting quantum information into classical information to facilitate subsequent wind speed prediction. L, M, and N represent the number of neurons in the quantum input layer, quantum hidden layer, and quantum output layer, respectively.

[0047] When input time vector When input to the quantum neural network, the quantum input layer converts input values ​​within the range [0,1] into quantum states with phase values ​​within the range [0,π / 2]. In this embodiment, the quantum input layer employs an angle data encoding strategy when converting classical data into quantum information carriers to efficiently and stably inject classical data into the quantum system. The output of the corresponding quantum input layer can be obtained through the following formula:

[0048]

[0049] In the formula, Here, x represents the output of the quantum output layer, and x represents the feature sequence value input to the quantum input layer. and It is the ground state representation symbol of a quantum bit. It is the orientation angle of the qubit in quantum space, where i is the imaginary unit. This is a mapping function, which can be defined by the user.

[0050] The quantum hidden layer employs quantum mechanics theory for feature transformation and entanglement learning, performing feature analysis in quantum space to predict subsequent wind speed sequences. The hidden layer simulates quantum interference effects through fully connected layers and activation functions in two complex neural networks. This effect is based on the principle that different complex probability amplitudes can cause constructive interference (signal enhancement) and destructive interference (signal cancellation). This characteristic cannot be directly simulated by pure real neural networks, making it particularly suitable for processing signals with phase and periodic information (such as time series like wind speed). The calculation process is shown in the following formula:

[0051]

[0052]

[0053] in and These are the complex weight matrix and complex weight bias of the complex neural network, respectively. and Similarly, It is a complex activation function. By simulating quantum interference effects, the model completes single-feature analysis and inference prediction of wind speed in the hidden layer.

[0054] The output signal of the hidden layer of the quantum-inspired neural network serves as the input data for the third layer (quantum output layer). The output layer's role is to map the analytical and predictive results of the quantum information on the single feature of wind speed in complex space back to the real number space that humans can read, facilitating observation. The above process consists of a complex fully connected neural network, an activation function, and a fully connected neural network. The calculation process is shown in the following formula:

[0055]

[0056]

[0057] in and These are the complex weight matrix and complex weight bias of the complex neural network, respectively. It is an activation function used to calculate the modulus of a complex number. It is a fully connected neural network used to linearly map vectors back to the output of the entire prediction model, which has a length of 24 and corresponds to the wind speed prediction for the next 24 hours.

[0058] It should be noted that, Figure 1 ①, ②, ③, etc. in the text represent the 24 outputs of the quantum neural network module, indicating that there are 24 outputs in a time sequence.

[0059] Step 3: Divide the data obtained in Step 1 into training and testing sets to train and test the short-term wind speed prediction model built in Step 2 accordingly.

[0060] Before training the model, as an optimization process, the model parameters of the short-term wind speed prediction model are first optimized using swarm intelligence algorithms, such as Particle Swarm Optimization (PSO) and Quantum Particle Swarm Optimization (QPSO).

[0061] As one implementation of the partitioning method, the data can be divided according to a certain ratio, such as a 7:3 ratio between the training and test sets. Alternatively, it can be divided into training, validation, and test sets according to a certain ratio. Furthermore, it should be noted that wind speed data exhibits significant time-series characteristics, requiring the dataset to be partitioned chronologically and a contextual prediction window to be set; random sampling cannot be used. Specifically, the preprocessed time series data is divided into two parts chronologically: the first 70% is used as the training set for learning and optimizing model parameters, and the remaining 30% is used as the test set to evaluate the generalization ability and prediction accuracy of the trained model.

[0062] The choice of loss function is crucial when training short-term wind speed forecasting models. It serves as the optimization objective for model learning and a benchmark for evaluating model performance. As an example of loss functions, two types are used in training short-term wind speed forecasting models: one measures the deviation between predicted and actual values, and the other measures the model's fit. The loss function directly affects the output at each time step, forcing the model's predictions to be as close as possible to the actual observations. The loss function measures the proportion of variance explained by the model, preventing local overfitting and focusing on the overall model's shape. In actual model training, a weighted sum combining both types of loss functions can be used to balance bias and variance, addressing the non-stationarity of wind speed sequences. Combining these two types of loss functions effectively improves the robustness and reliability of short-term wind speed forecasting models under complex weather conditions.

[0063] Among these, as a preferred processing method, the loss function used to measure the deviation between the predicted and actual values ​​can specifically be the mean absolute error (MAE), root mean square error (RMSE), or mean absolute percentage error (MAPE). Let... , Let and y represent the predicted wind speed, the average wind speed, and the actual wind speed, respectively. Let K be the number of statistical data points studied, k = 1, 2, ..., K. The corresponding formulas for MAE, RMSE, MAPE, and R2-Score are as follows:

[0064]

[0065]

[0066]

[0067]

[0068]

[0069]

[0070] Specifically, the training process is as follows: Figure 3As shown, the acquired raw wind speed time series dataset is divided into four datasets according to season. Each seasonal dataset is further divided into training, validation, and test sets, thus obtaining short-term wind speed prediction models corresponding to each season; that is, one short-term wind speed prediction model corresponds to one season. 80% of the data is used as the training set, and 20% is used as the validation and test sets, meaning the validation and test sets use the same batch of data. Ensuring the diversity and representativeness of the dataset guarantees the model's generalization ability. The epochs (number of iterations / training rounds) and batch size are adjusted to 10000 and 500, respectively. The Adam optimizer (Adaptive Moment Estimation) is used to update the model parameters. The Adam optimizer effectively avoids the problem of excessively high or low learning rates that may occur in traditional Stochastic Gradient Descent (SGD) by adaptively adjusting the learning rate. A learning rate warm-up strategy is employed to improve training stability, accelerate model convergence, and enhance model performance. A learning rate decay strategy is adopted to automatically adjust the learning rate based on the preset number of epochs or performance changes during training, thereby maintaining the stability of the training process.

[0071] Step 4: Use the trained short-term wind speed prediction model to make short-term wind speed predictions.

[0072] The process involves acquiring the time-series data required for short-term wind speed forecasting and inputting it into a pre-trained short-term wind speed forecasting model corresponding to the season in which the time-series data belongs, thereby obtaining the short-term wind speed forecast result. The specific data required for short-term wind speed forecasting depends on the training data. As described in step one, one embodiment of the data required for short-term wind speed forecasting includes historical meteorological characteristic time-series data, including wind speed, air pressure, atmospheric temperature, and oxygen concentration. For example, when the time-series data required for short-term wind speed forecasting is summer data, it is input into the short-term wind speed forecasting model trained using summer data.

[0073] Experiments were conducted to verify the effectiveness of the method of the present invention. Figure 4The experimental results of the proposed method are presented. The red line (Real) represents the actual wind speed value, the green line (Proposed method) represents the result of the short-term wind speed prediction model based on deep learning and quantum neural networks, and the blue line (Cnn-LSTM-transformer) represents the prediction result of the comparison model (Cnn-LSTM-transformer). The comparison model (Cnn-LSTM-transformer) is a model where CNN, LSTM, and Transformer networks are directly connected in series. The results show that the short-term wind speed prediction model based on deep learning and quantum neural networks of this invention clearly provides predictions that are closer to the actual results and are more accurate.

[0074] An implementation method for a short-term wind speed prediction system based on deep learning and quantum neural networks:

[0075] A short-term wind speed prediction system based on deep learning and quantum neural networks according to the present invention is illustrated in the figure below. Figure 5 As shown, the system includes a memory, a processor, a system bus, and a computer program stored in the memory. The processor and memory communicate and interact with each other via the system bus. The processor executes the computer program to implement the steps of the short-term wind speed prediction method based on deep learning and quantum neural networks of this invention. The processor can be a microprocessor (MCU) or other processing device; the memory can be any type of memory that stores information using electrical energy, such as non-volatile storage media (including computer programs and databases), or other types of memory.

[0076] The system offers high deployment flexibility. It can be deployed on cloud servers, leveraging their powerful computing and storage resources to process large-scale meteorological data and enable forecasting and remote operation and maintenance across multiple wind farm clusters. Alternatively, it can be deployed on edge computing nodes or industrial control computers at wind farm sites to achieve local real-time data processing and low-latency forecasting, reducing reliance on network bandwidth and ensuring data privacy and security. Furthermore, the system can be provided as a software-as-a-service, allowing users to access forecasting functions on demand via terminals, effectively reducing hardware investment and maintenance costs at wind farms.

[0077] One embodiment of a computer-readable storage medium:

[0078] This invention provides a computer-readable storage medium storing a computer program. The computer program is executed by a processor to implement the steps of the short-term wind speed prediction method based on deep learning and quantum neural networks as described above. The specific short-term wind speed prediction method based on deep learning and quantum neural networks has been described in sufficient detail in the above embodiments and will not be repeated here.

[0079] In summary, the present invention has the following characteristics:

[0080] First, a model for short-term wind speed prediction based on deep learning and quantum neural networks is proposed. This aims to improve the accuracy and generalization of wind speed prediction, and address issues such as low prediction efficiency, high computational resource consumption, and low reliability. Second, experiments yield the following conclusions: On the one hand, by introducing deep learning theory, the global search capability and the ability to escape local optima are greatly enhanced during training; its main advantages can be summarized as stronger globality and more stable performance. On the other hand, quantum neural networks, utilizing quantum superposition and parallelism, can achieve exponential high-dimensional data processing and computational acceleration with very few parameters, representing a key potential for future computing; they also demonstrate strong potential in capturing the complex spatiotemporal dynamic characteristics that govern wind speed fluctuations. Finally, by optimizing the loss function and learning rate, the model converges faster during training. Comparative experimental results show that the model designed in this invention exhibits significant advantages in the field of wind speed prediction compared to other traditional prediction methods.

Claims

1. A short-term wind speed prediction method based on deep learning and quantum neural networks, characterized in that, The method includes: Obtain the time series data required for short-term wind speed prediction and input it into a trained short-term wind speed prediction model corresponding to the season to which the time series data belongs, and obtain the short-term wind speed prediction result. The short-term wind speed prediction model includes a feature extraction and alignment module and a wind speed prediction module. The feature extraction and alignment module is used to extract features from the data input to the model using a deep learning network and then align the features so that the current tensor is mapped to the quantum neural network space. The wind speed prediction module is used to predict the wind speed using the aligned features using a quantum neural network.

2. The short-term wind speed prediction method based on deep learning and quantum neural networks according to claim 1, characterized in that, The parts of a deep learning network used for feature extraction include convolutional neural network layers and fusion network layers. The fusion network layers are used to flatten multidimensional inputs and then fuse features.

3. The short-term wind speed prediction method based on deep learning and quantum neural networks according to claim 1, characterized in that, In deep learning networks, the part used for feature alignment is the gated recurrent neural network layer.

4. The short-term wind speed prediction method based on deep learning and quantum neural networks according to claim 1, characterized in that, The required time series data includes historical meteorological feature time series data.

5. The short-term wind speed prediction method based on deep learning and quantum neural networks according to claim 4, characterized in that, Meteorological characteristics include wind speed, air pressure, atmospheric temperature, and oxygen concentration.

6. The short-term wind speed prediction method based on deep learning and quantum neural networks according to claim 1, characterized in that, The quantum input layer in a quantum neural network employs an angle data encoding strategy to convert classical input information into a quantum information carrier.

7. The short-term wind speed prediction method based on deep learning and quantum neural networks according to claim 1, characterized in that, The quantum output layer in a quantum neural network employs a quantum amplitude sampling strategy to convert quantum information into classical information for wind speed prediction.

8. The short-term wind speed prediction method based on deep learning and quantum neural networks according to any one of claims 1 to 7, characterized in that, The loss function for training a short-term wind speed prediction model includes a loss function that measures the deviation between the predicted value and the actual value, and a loss function that measures the goodness of fit of the model.

9. A short-term wind speed forecasting system based on deep learning and quantum neural networks, comprising a processor, characterized in that, The processor executes a computer program to implement the steps of the method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 8.

Images