A short-term load prediction method based on quantum hybrid features

Abstract

This invention claims protection for a short-term load forecasting method based on quantum hybrid features, belonging to the field of smart grid technology. The method includes: acquiring historical load data and extracting multiple classical features, including hysteresis features; inputting the classical features into a pre-constructed parameterized quantum circuit to generate four quantum features through measurement; fusing the classical and quantum features to form a hybrid feature vector; and inputting the hybrid feature vector into a pre-trained XGBoost regression model to obtain the short-term load forecast result. Furthermore, the SHAP algorithm is used to perform interpretability analysis on the forecast result, quantifying the contribution of each feature. This invention improves forecast accuracy and robustness under extreme conditions by fusing classical time-series features and quantum features. Simultaneously, SHAP analysis solves the black-box problem of the model, enhancing the model's credibility in critical power infrastructure, and possesses the advantages of high accuracy, strong robustness, and high interpretability.

Description

Technical Field

[0001] This invention belongs to the field of smart grid technology, specifically to a short-term load forecasting method based on quantum hybrid characteristics. Background Technology

[0002] Short-term load forecasting is a crucial foundation for smart grid dispatching and operation. Accurate load forecasting can effectively optimize generation plans, reduce reserve capacity, and improve the economy and security of grid operation. However, power load data exhibits strong nonlinearity, time-varying characteristics, and complex periodicity, posing significant challenges to the forecasting process. Existing technologies for short-term load forecasting mainly include traditional time series analysis and data-driven machine learning methods. Traditional methods, such as the ARIMA model, while structurally simple, struggle to capture the complex nonlinear relationships within load data, resulting in limited forecasting accuracy.

[0003] In recent years, data-driven machine learning methods have become mainstream. One type of method is based on recurrent neural networks and their variants. For example, CN121602346A discloses a weather-driven active distribution network load forecasting method, which constructs weather-derived features and uses a bidirectional long short-term memory network (BiLSTM) combined with an attention mechanism for forecasting. However, this type of method heavily relies on the quality of weather features, and its prediction accuracy and robustness are often difficult to guarantee in scenarios where data distribution shifts due to extreme weather or other conditions. Another type of method focuses on improving model performance through feature engineering. For example, CN121546560A proposes a short-term load forecasting method based on feature selection, which selects key features through feature importance analysis and maximum information coefficient before inputting them into the prediction model. However, the features selected by this type of method are still limited to classic time-series features or weather features. Essentially, it is an optimization of existing feature dimensions and does not create new feature dimensions that can represent complex nonlinear patterns, thus creating a bottleneck for improving prediction performance.

[0004] In addition, some methods attempt to incorporate physical information or perform scenario classification to enhance predictive capabilities. For example, CN121328859A discloses a short-term load forecasting method for microgrids, embedding physical laws such as power balance into an LSTM model in the form of a loss function. CN121390474A proposes a short-term load forecasting method for large-scale distribution areas, which uses iTransformer for forecasting after classifying distribution areas by temperature sensitivity and correcting temperature characteristics. While these methods have their innovations, they still fail to address the problem of insufficient robustness of the models under extreme conditions, and their decision-making processes often lack transparency, resulting in poor model interpretability. This makes it difficult for dispatchers to trust and intervene in the model's predictions at critical moments, limiting their widespread application in critical power infrastructure.

[0005] Therefore, there is an urgent need for a short-term load forecasting method that can overcome the limitations of existing feature dimensions, and has high accuracy, strong robustness, and good interpretability, in order to overcome one or more of the defects in the existing technologies. Summary of the Invention

[0006] This invention aims to solve the problems of the prior art. It proposes a short-term load forecasting method and system based on quantum hybrid characteristics, which features high prediction accuracy, strong robustness, and good interpretability. The technical solution of this invention is as follows:

[0007] A short-term load prediction method based on quantum hybridity features includes the following steps:

[0008] Step 1: Obtain historical power load data for the target area and extract a classic feature set from the historical data; wherein, the classic feature set includes at least one lagging load feature;

[0009] Step 2: Input the classical feature set into a pre-constructed parameterized quantum circuit to generate a quantum feature set; wherein, the parameterized quantum circuit includes 4 qubits for constructing quantum basic units and multiple entanglement layers for establishing associations for the qubits, for mapping classical features to a high-dimensional quantum feature space;

[0010] Step 3: Fuse the classical feature set and the quantum feature set to construct a hybrid feature vector;

[0011] Step 4: Input the hybrid feature vector into the pre-trained gradient boosting tree-based regression model for processing, and output the short-term load prediction result;

[0012] Step 5: Based on the Shapley additive interpretation algorithm, calculate the marginal contribution of each feature in the mixed feature vector to the short-term load forecast result, and generate interpretability analysis results.

[0013] Furthermore, the classic feature set of step 1 includes: a first lag feature, which is the load value one hour before the prediction point; a second lag feature, which is the load value 24 hours before the prediction point; a third lag feature, which is the load value 168 hours before the prediction point; and a differential feature, which is the load difference between one hour before the prediction point and the hour before that.

[0014] Furthermore, the parameterized quantum circuit in step 2 is a 4-qubit circuit, including a multi-layer data re-upload structure. Each layer includes a rotating gate layer composed of single-qubit rotating gates and an entangled gate layer composed of multi-qubit controlled gates. The generation of the quantum feature set includes measuring the expected values ​​of the Pauli X, Y, and Z operators for each qubit to generate a 12-dimensional quantum feature set.

[0015] Furthermore, step 2 inputs the classical feature set into the pre-constructed parameterized quantum circuit, including repeatedly uploading the classical feature set as an encoding angle to the multi-layer structure of the parameterized quantum circuit.

[0016] Furthermore, the gradient boosting tree model in step 4 is either the XGBoost model or the LightGBM model.

[0017] Furthermore, before constructing the hybrid feature vector in step 3, robust normalization processing is performed on the classical feature set and the quantum feature set respectively.

[0018] Furthermore, the method also includes a model training step: constructing an initial gradient boosting tree model; training and early stopping control of the initial gradient boosting tree model using a training set and a validation set to obtain the pre-trained gradient boosting tree model.

[0019] Furthermore, the model training step also includes incremental learning on the training set using a new dataset and monitoring performance on the validation set to fine-tune the pre-trained gradient boosting tree model to adapt to the new load prediction scenario.

[0020] Furthermore, the output of short-term load forecast results includes traversing the test set using a non-overlapping sliding window, constructing inputs based on the real observations of the lag characteristics within the window, and outputting independent forecast values ​​corresponding to the window.

[0021] A short-term load prediction system based on quantum hybrid characteristics, comprising:

[0022] A classic feature extraction module is used to acquire historical power load data of a target area and extract a classic feature set from the historical data; wherein the classic feature set includes at least one lagging load feature;

[0023] A quantum feature generation module is used to input the classical feature set into a pre-constructed parameterized quantum circuit to generate a quantum feature set; wherein, the parameterized quantum circuit contains multiple qubits and multiple entanglement layers, used to map classical features to a high-dimensional quantum feature space;

[0024] The feature fusion module is used to fuse the classical feature set and the quantum feature set to construct a hybrid feature vector;

[0025] The load prediction module is used to input the hybrid feature vector into a pre-trained gradient boosting tree model for processing and output short-term load prediction results.

[0026] The interpretability analysis module is used to calculate the marginal contribution of each feature in the hybrid feature vector to the short-term load forecast result based on the SHAP algorithm, and generate interpretability analysis results.

[0027] The advantages and beneficial effects of this invention are as follows:

[0028] 1. High prediction accuracy: The quantum features generated by parameterized quantum circuits effectively expand the dimension of the feature space, capturing complex nonlinear correlations that are difficult to detect with classical features, and providing richer information for gradient boosting tree models. On the test set, the coefficient of determination R² of this invention reaches 0.9991, and the mean absolute error (MAE) is reduced to 0.3598 MW, significantly outperforming existing technologies.

[0029] 2. Strong robustness: SHAP analysis shows that although quantum features contribute less under standard conditions, they play a crucial role in ensuring robustness in complex scenarios such as extreme weather. This ensures that the model's performance is almost unaffected when the data distribution shifts, thus solving the problem of insufficient robustness of existing technologies such as CN121602346A under extreme conditions.

[0030] 3. Transparent and highly interpretable decision-making process: By introducing the SHAP algorithm, the marginal contribution of each classical and quantum feature to the prediction results can be accurately quantified, generating clear and interpretable analysis results. This makes the model's decision-making logic completely transparent, solving the problem of opaque decision-making processes in existing technology models and enhancing its credibility in critical power infrastructure.

[0031] 4. Good generalization ability: Through an incremental learning strategy of fine-tuning on new datasets, the model can quickly adapt to the load characteristics of different regions and seasons, and has achieved excellent prediction results on multiple public datasets such as ETT and Ausgrid, proving its good cross-domain generalization ability.

[0032] 5. Innovation Analysis: According to claims 1-4, this invention employs a quantum circuit with fixed parameters for feature extraction, which differs from existing quantum machine learning schemes that require end-to-end training, effectively avoiding instability issues during training. Furthermore, SHAP analysis reveals a complementary mechanism by which quantum features play a crucial robust role under extreme weather conditions, a mechanism that differs from the conventional understanding that feature contribution is positively correlated with prediction accuracy. Attached Figure Description

[0033] Figure 1 This is a flowchart illustrating a preferred embodiment of the short-term load forecasting method based on quantum hybrid features provided by the present invention.

[0034] Figure 2This is a comparison chart of the predicted and actual values ​​for 168 hours under non-extreme and extreme weather conditions in an embodiment of the present invention.

[0035] Figure 3 This is a summary diagram of the importance of SHAP features for the E1 and E2 models in this embodiment of the invention.

[0036] Figure 4 This is a complete 168-hour prediction graph of the ETT dataset in this embodiment of the invention;

[0037] Figure 5 This is a comparison chart of the error distribution in the generalization experiment of the Australian Ausgrid dataset in this embodiment of the invention. Detailed Implementation

[0038] The technical solutions of the embodiments of the present invention will be clearly and thoroughly described below with reference to the accompanying drawings. The described embodiments are merely some embodiments of the present invention.

[0039] The technical solution of the present invention to solve the above-mentioned technical problems is:

[0040] Example 1

[0041] This embodiment provides a short-term load forecasting method based on quantum hybrid characteristics. For example... Figure 1 As shown, the method includes the following steps:

[0042] 1. Data Acquisition and Preprocessing

[0043] Historical power load data for the target area was obtained. In this embodiment, multiple datasets were used for validation, including transformer load data from Guangxi, China, the ETT dataset, and the Ausgrid dataset from Australia. The data sampling frequency was once per hour.

[0044] Preprocessing of the raw data includes:

[0045] Data cleaning: Time series interpolation was used to fill in missing values, and box plots were used to identify and remove outliers.

[0046] Feature selection: Core classic features are selected from the original 39 candidate features through variance inflation factor filtering and Granger causality test. In a preferred embodiment, the selected classic feature set includes 4 features, as described in step 2.

[0047] Data standardization: To eliminate the influence of different feature units and improve the robustness of the model, RobustScaler is used to process the feature data. This scaler uses the median and quartiles and is insensitive to outliers. Standardized parameters are fitted on the training set and then applied to the validation and test sets.

[0048] Dataset partitioning: Strictly following the principles of time series forecasting, the data was divided into training, validation, and test sets in a 70:15:15 ratio.

[0049] 2. Classic Feature Set

[0050] Classic feature sets are extracted from preprocessed historical data. In a preferred embodiment, the classic feature set includes the following four features:

[0051] LOAD_LAG_1h: Load value one hour before the forecast point, used to capture real-time load trends.

[0052] LOAD_LAG_24h: Load value 24 hours before the forecast point, used to capture the daily periodicity of the load.

[0053] LOAD_LAG_168h: Load value 168 hours before the forecast point (i.e., the same time in the previous week), used to capture the periodicity of the load.

[0054] LOAD_DIFF_1H: The load difference between the hour before and the hour before the forecast point, used to capture short-term load fluctuations.

[0055] Let the input sample at time t be... It consists of the four features mentioned above.

[0056] 3. Generate quantum feature sets

[0057] The classical feature set obtained in step 2 The input is fed into a pre-constructed parameterized quantum circuit to generate a set of quantum features. The parameterized quantum circuit structure used in this embodiment is shown below. Figure 1 The Quantum-Inspired Feature Generator, detailed below:

[0058] Number of qubits: 4 qubits are used.

[0059] Circuit layers: A 3-layer data re-upload structure is used. Each layer has the same structure.

[0060] Data encoding: For each layer Classic features The rotation angle is sequentially encoded onto each qubit. The encoding gates are combined as follows: Door and Gate. For a qubit q, the operation at the l-th layer can be represented as: Because it has a 3-layer structure, the same 4 feature values ​​will be uploaded 3 times.

[0061] Entanglement layer: After the data encoding gate in each layer, a CNOT gate is applied to form an entanglement layer, establishing quantum entanglement between different qubits, thereby capturing the correlation between features. This embodiment can employ a linear chain or a ring-connected topology.

[0062] Quantum state preparation: The initial state of the entire circuit can be set as a uniform superposition state. To avoid initialization deviations.

[0063] Measurement: After the circuit processing described above, the expected value of the Pauli operator is measured for each qubit. Specifically, for each qubit q, the expected values ​​of the Pauli-X, Y, and Z operators are measured respectively, thereby generating a 12-dimensional quantum eigenvector. The value range of each feature is [−1, 1].

[0064] The entire feature generation process can be represented by the following formula. First, the input x is encoded into a quantum state:

[0065]

[0066] in, For the first Layer The encoding angle of each qubit.

[0067] Then, measure the expected value of the Pauli operator:

[0068]

[0069]

[0070]

[0071] 4. Constructing a hybrid feature vector

[0072] The four classical features obtained in step 2 and the twelve quantum features obtained in step 3 are concatenated to form a 16-dimensional hybrid feature vector. :

[0073]

[0074] 5. Load forecasting

[0075] The hybrid feature vector constructed in step 4 The input is fed into a pre-trained gradient boosting tree model for processing, and the output is a short-term load forecast result.

[0076] Model Selection: In this embodiment, the gradient boosting tree model is preferably XGBoost or LightGBM. These models perform well on tabular data prediction tasks and are easy to combine with the SHAP algorithm for interpretability analysis.

[0077] Model training:

[0078] Pre-training: The XGBoost model is pre-trained using the training set. The model's objective function is optimized for the regression task and includes squared loss and regularization terms to prevent overfitting. The objective function can be expressed as:

[0079]

[0080] in, This represents the overall objective function of the model with ϕ as a parameter. For the squared loss term, This is a regularization term.

[0081] Early stopping: During training, an early stopping strategy is implemented using the validation set. Training is stopped when the loss on the validation set no longer decreases over several consecutive training epochs to prevent overfitting.

[0082] Fine-tuning: When it is necessary to apply the model to a new dataset, a fine-tuning strategy can be adopted. Starting with the pre-trained model, incremental learning is performed on the training set of the new dataset, combined with early stopping on the validation set, so that the model can quickly adapt to the load characteristics of the new domain while retaining the original knowledge.

[0083] Prediction: During the testing phase, a non-overlapping sliding window is used to traverse the test set, and the input is constructed based on the true observations of the lag features within the window. It outputs the independent predicted values ​​corresponding to the window. Ensure that the forecast is strictly forward-looking and avoid the propagation of errors.

[0084] 6. Interpretability Analysis

[0085] To explain the model's prediction logic, the SHAP algorithm is introduced. Based on the concept of Shapley value in game theory, SHAP calculates a weighted average of the marginal contributions of each feature to all possible feature subsets. For a specific prediction sample, the SHAP value of the i-th feature is... The calculation formula is:

[0086]

[0087] in, Let be the SHAP value of the i-th feature, F be the set of all features, and S be the subset of features that does not include feature i. The number of features of subset S. The predicted output of the model when using features in S. This is the model's predicted output after adding i to S.

[0088] SHAP analysis can clearly reveal the decision-making basis of the model. For example, in this invention, SHAP analysis can prove that classical features, especially LOAD_LAG_1h, play a dominant role in prediction accuracy, accounting for 91.5% of the total contribution; the nonlinear contribution mode of quantum features disperses the risk of model dependence on a single classical feature, thus providing crucial robustness assurance in complex scenarios such as extreme weather.

[0089] To verify the technical effectiveness of this invention, experiments were conducted using transformer load data from Guangxi, China, the ETT dataset, and the Ausgrid dataset from Australia.

[0090] Prediction accuracy: As shown in Table 1, the RX model using the method of this invention achieves an R² of 0.9991 and an MAE of 0.3598 MW on the Guangxi dataset test set. It uses the mainstream algorithm of the RX model and has considerable prediction accuracy.

[0091] Table 1 Comparison of Load Forecasting Results

[0092]

[0093] Ablation Experiments and Robustness: As shown in Tables 2 and 3, the E1 and E2 models were compared through ablation experiments. The E1 model used only classical features, while the E2 model used both classical and quantum features. On the standard test set, the MAE of the E1 model was slightly lower than that of the E2 model, indicating that quantum features introduced some noise under standard conditions. However, on the extreme weather subset, the R² fluctuation of the E2 model was much smaller than that of the E1 model, demonstrating the crucial role of quantum features in ensuring model stability under extreme conditions. Figure 2 The model's prediction results are shown on both the non-extreme weather dataset and the extreme weather dataset in Guangxi.

[0094] Table 2 Ablation Experiment Results

[0095]

[0096] Table 3 Comparison of performance under extreme and non-extreme weather conditions

[0097]

[0098] Interpretability analysis: such as Figure 3 As shown, SHAP analysis clearly reveals a decision-making mechanism dominated by classical features and complemented by quantum features. Classical features are the core of prediction accuracy, while quantum features, through their nonlinear contribution patterns, enhance the model's robustness in complex scenarios.

[0099] Generalization ability: such as Figure 4 As shown, through fine-tuning, the model achieves R² of 0.9522 on the ETT dataset. Figure 5 As shown, the average R² reaches 0.9722 on the Australian Ausgrid dataset, demonstrating that the method of this invention has good cross-domain generalization ability.

[0100] These experimental results fully demonstrate that, compared with existing technologies such as CN121602346A, CN121546560A, CN121328859A, and CN121390474A, the present invention has achieved certain technical progress in terms of prediction accuracy, robustness, and interpretability.

[0101] It should be noted that the user information (including but not limited to user device information, personal user information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the laws, regulations and standards of relevant countries and regions.

[0102] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions.

[0103] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0104] The above embodiments should be understood as illustrative only and not as limiting the scope of protection of the present invention. After reading the description of the present invention, those skilled in the art can make various alterations or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.

Claims

1. A short-term load forecasting method based on quantum hybrid features, characterized in that, Includes the following steps: Step 1: Obtain historical power load data for the target area and extract a classic feature set from the historical data; wherein, the classic feature set includes at least one lagging load feature; Step 2: Input the classical feature set into a pre-constructed parameterized quantum circuit to generate a quantum feature set; wherein, the parameterized quantum circuit includes 4 qubits for constructing quantum basic units and multiple entanglement layers for establishing associations for the qubits, for mapping classical features to a high-dimensional quantum feature space; Step 3: Fuse the classical feature set and the quantum feature set to construct a hybrid feature vector; Step 4: Input the hybrid feature vector into the pre-trained gradient boosting tree-based regression model for processing, and output the short-term load prediction result; Step 5: Based on the Shapley additive interpretation algorithm, calculate the marginal contribution of each feature in the mixed feature vector to the short-term load forecast result, and generate interpretability analysis results.

2. The method of claim 1, wherein, The classic feature set in step 1 includes: a first lag feature, which is the load value one hour before the prediction point; a second lag feature, which is the load value 24 hours before the prediction point; a third lag feature, which is the load value 168 hours before the prediction point; and a differential feature, which is the load difference between one hour before the prediction point and the hour before that.

3. The method of claim 1, wherein, The parameterized quantum circuit in step 2 is a 4-qubit circuit, including a multi-layer data re-upload structure. Each layer includes a rotating gate layer composed of single-qubit rotating gates and an entangled gate layer composed of multi-qubit controlled gates. The generation of the quantum feature set includes measuring the expected values ​​of the Pauli X, Y, and Z operators for each qubit to generate a 12-dimensional quantum feature set.

4. The method of claim 3, wherein, Step 2 involves inputting the classical feature set into a pre-constructed parameterized quantum circuit, including repeatedly uploading the classical feature set as an encoding angle to the multi-layer structure of the parameterized quantum circuit.

5. The method of claim 1, wherein, The gradient boosting tree model in step 4 is either the XGBoost model or the LightGBM model.

6. The method according to claim 1, characterized in that, Before constructing the hybrid feature vector in step 3, robust normalization processing is performed on the classical feature set and the quantum feature set respectively.

7. The method of claim 1, wherein, The method further includes a model training step: constructing an initial gradient boosting tree model; training the initial gradient boosting tree model and performing early stopping control using a training set and a validation set to obtain the pre-trained gradient boosting tree model.

8. The method of claim 7, wherein, The model training steps also include incremental learning on the training set using the new dataset and monitoring performance on the validation set to fine-tune the pre-trained gradient boosting tree model to adapt to the new load prediction scenario.

9. The method of claim 1, wherein, The output short-term load forecast results include traversing the test set using a non-overlapping sliding window, constructing inputs based on the real observations of the lag characteristics within the window, and outputting independent forecast values ​​corresponding to the window.

10. A short-term load forecasting system based on quantum hybrid features, characterized by, include: A classic feature extraction module is used to acquire historical power load data of a target area and extract a classic feature set from the historical data; wherein the classic feature set includes at least one lagging load feature; A quantum feature generation module is used to input the classical feature set into a pre-constructed parameterized quantum circuit to generate a quantum feature set; wherein, the parameterized quantum circuit contains multiple qubits and multiple entanglement layers, used to map classical features to a high-dimensional quantum feature space; The feature fusion module is used to fuse the classical feature set and the quantum feature set to construct a hybrid feature vector; The load prediction module is used to input the hybrid feature vector into a pre-trained gradient boosting tree model for processing and output short-term load prediction results. The interpretability analysis module is used to calculate the marginal contribution of each feature in the hybrid feature vector to the short-term load forecast result based on the SHAP algorithm, and generate interpretability analysis results.

Images