Hierarchical explainable analysis method and system for power load forecasting model

Abstract

The application discloses a hierarchical explainable analysis method and system for a power load prediction model, and comprises the following steps: collecting power load historical data containing multi-dimensional features, performing multi-source fusion and preprocessing to construct a load historical data set; inputting the data set into a pre-trained model to generate a load prediction sequence; decomposing the historical load data into trend, seasonal and dynamic residual components; calculating the contribution of each component to the prediction to obtain a concept-level explanation; extracting a local morphological prototype in a time series feature space and reversely mapping it to an original feature space to realize morphological-level and physical-level explanations by correlating physical variables; and cross-verification of the three-level explanation results to generate a comprehensive explainable analysis conclusion. The application realizes real-time load prediction and explanation through an online system. The method can realize cross-scale and hierarchical model explanation, improve prediction transparency and understandability, and provide a reliable decision basis for power dispatching and planning.

Description

Technical Field

[0001] This invention belongs to the field of interpretable artificial intelligence, specifically relating to a hierarchical interpretable analysis method and system for power load forecasting models. Background Technology

[0002] With the increasing integration of renewable energy into the grid and the diversification of demand-side electricity consumption, the non-stationarity and random volatility of power load sequences continue to intensify. In practical dispatching decisions, improved forecast accuracy alone cannot fully meet the needs of dispatchers. Faced with load fluctuations caused by extreme weather, holidays, or adjustments in energy policies, dispatchers not only need to know the forecast values ​​but also understand the driving factors behind the forecast results. Therefore, improving the interpretability of load forecasting models is of practical significance for enhancing the reliability of these models in engineering applications.

[0003] While existing deep learning prediction methods have achieved good accuracy, they lack transparency. Current model interpretability methods have made some progress. Intrinsically interpretable methods achieve direct explanation of the prediction process by constructing relatively simple and transparent models. For example, linear regression models characterize the marginal impact of input features on the output through parameter coefficients; decision tree models explicitly express the decision path by dividing the feature space layer by layer through a tree-like rule structure. Posteriorly interpretable methods primarily target black-box models, achieving ex-post explanation of prediction results through analysis of input perturbations or internal model responses. These methods do not change the model structure but analyze model behavior after training. Locally Interpretable Model-agnostic Explanations (LIME) achieve approximate explanations of single-sample prediction results by constructing interpretable surrogate models in the local neighborhood of the sample; the Shapley value-based SHAP method calculates the marginal contribution of features to the prediction results from a cooperative game theory perspective, thus achieving a unified attribution framework. The Shapelet method achieves explicit characterization of the local structure of time series by mining key subsequences with discriminative power, and demonstrates good interpretability in classification tasks. Subsequently, the Prototype Network learns a set of representative prototype vectors in the feature space and uses the distance between samples and prototypes as the decision criterion, providing a similarity-based explanation mechanism for the model.

[0004] While the aforementioned methods have achieved some success in power load forecasting, they still have several shortcomings when facing the intuitive interpretability requirements of this field. First, regarding the granularity of interpretation, existing methods mostly focus on point-level feature contribution analysis, lacking the ability to uniformly model the relationship between macroscopic physical components and microscopic local morphology, thus making it difficult to form a hierarchical interpretation structure across scales. Second, regarding the consistency of interpretation in regression tasks, due to the significant non-stationarity of power load data, existing attribution methods based on perturbations or gradients are prone to fluctuations under different time distributions, affecting the consistency between interpretation results and model behavior. Furthermore, in terms of physical semantic expression, current interpretation results are mostly presented in the form of numerical weights or importance scores, lacking a direct correspondence with engineering phenomena such as load trend changes and peak-valley structures, thus limiting their application value in power dispatching and planning decisions.

[0005] Current load forecasting models and interpretable methods face the following challenges: (1) The point-level interpretation granularity is low. Existing interpretation methods are usually limited to analyzing the point-level contribution of a single feature, failing to effectively combine macroscopic physical components and microscopic local morphology for unified modeling. Due to the lack of a hierarchical interpretation structure across scales, these methods struggle to capture the complex relationships between different levels in time series data, resulting in limitations on the accuracy and interpretability of the interpretation results. To gain a more comprehensive understanding of model behavior, multi-level fusion at the interpretation granularity is necessary to provide a richer analytical framework.

[0006] (2) Insufficient consistency in interpretation for load forecasting tasks. In regression tasks such as power load forecasting, due to the non-stationary nature of the data, existing interpretation methods based on perturbations or gradients are prone to fluctuations under different time distributions, which leads to the disruption of consistency between the interpretation results and the actual model behavior. Power load data has strong seasonality, trends, and random fluctuations. These factors make the attribution results generated by traditional interpretation methods unstable in different time periods, thus affecting the reliability and practicality of the interpretation.

[0007] (3) Insufficient physical semantic expression. Most current interpretation methods demonstrate the influence of features through numerical weights or importance scores. While these results can be quantified, they lack a direct connection with actual engineering phenomena of power load, such as load trend changes and peak-valley structure. Therefore, existing interpretation methods are difficult to provide intuitive and easily understandable basis for power dispatching and planning decisions, limiting their value and operability in practical applications. It is necessary to enhance the physical semantic expression to directly correspond the model output with the physical phenomena of the actual power system, thereby improving its application effect in real-world scenarios. Summary of the Invention

[0008] In view of this, to address the problems of low granularity, poor consistency of interpretation results, and poor physical semantics faced by existing interpretation methods in power load forecasting, a three-layer interpretation system of "component contribution analysis—morphological pattern extraction—physical semantic alignment" is adopted to analyze the decision-making process of load forecasting from macro to micro levels, providing intuitive reference for power grid dispatchers. This method not only improves the transparency of deep model predictions but also significantly enhances the interpretability of load forecasting models and strengthens their credibility in engineering applications, possessing significant theoretical and practical value.

[0009] To solve the above-mentioned technical problems, the present invention provides the following technical solution: A hierarchical interpretable analysis method for power load forecasting models includes the following steps: Collect historical power load data containing multidimensional features, and perform multi-source data fusion and preprocessing to construct a historical load dataset; Input the historical load dataset into the trained power load prediction model, obtain the load prediction sequence output by the power load prediction model, and define the feature space composed of the multidimensional features as the original feature space. The historical power load data is decomposed to obtain trend components, seasonal components, and dynamic residual components; By perturbing or reconstructing the inputs of the trend component, seasonal component, and dynamic residual component, the contribution value of each component in the predicted output is calculated, and a conceptual interpretation result is obtained. Local morphological prototypes with discriminative capabilities are extracted from the time series feature space corresponding to the load prediction sequence to obtain morphological interpretation results; The local morphological prototype is mapped to the original feature space. Based on the mapping result, the local morphological prototype is correlated with external physical variables to determine the dominant physical factors that cause the local morphology and obtain physical-level interpretation results. By combining the conceptual, morphological, and physical interpretation results, multi-level interpretable analysis results of the power load prediction model are generated. Based on the results of the multi-level interpretable analysis, a hierarchical interpretation process for load forecasting is constructed, and the hierarchical interpretation process is used to perform real-time interpretation and analysis of online load forecasts.

[0010] Furthermore, the multidimensional features include time features, weather features, and social scheduling features, and the construction of the load dataset specifically includes: Collect historical power load data with multidimensional features, including time features such as year, month, day, hour, and minute; weather features such as temperature, humidity, and wind speed; and social dispatch features such as electricity price and holidays. Multi-source data fusion and preprocessing are performed on historical power load data from different data sources to construct a historical load dataset; The historical load dataset is divided into training set, validation set and test set.

[0011] Furthermore, obtaining the load forecast sequence output by the power load forecasting model includes: A deep load forecasting model is selected, trained based on the training set, with mean squared error as the loss function, and the effect is verified using a validation set. The trained model is then saved as the power load forecasting model. The test set is predicted based on the power load prediction model to obtain the load prediction sequence, and the prediction effect of the power load prediction model is tested.

[0012] Furthermore, the decomposition of the historical power load data includes: The historical power load data input into the power load prediction model is decomposed using a seasonal trend decomposition method based on local weighted regression to obtain trend components, seasonal components, and residual components. Feature dimension pooling is performed on the trend component, seasonal component, and residual component to reduce the feature dimension.

[0013] Furthermore, the contribution of each component to the predicted output is calculated, yielding conceptual-level explanations including: The trend component, seasonal component, and residual component are defined as three types of conceptual participants, forming a set. ,in These are the concepts of trend component, seasonal component, and residual component, respectively, with a set size of m=3; For a given input sample of the power load forecasting model to be interpreted , No. i One concept Contribution value The calculation formula is as follows:

[0014] in, It does not include All possible subsets of concepts It indicates the combination of concepts. The predicted output of the already trained power load forecasting model, i.e., incorporating concepts. The subsequent prediction results; In the combination of concepts The predicted output of the already trained power load forecasting model, i.e., without considering the concept. At that time, the prediction result is constructed by using a baseline replacement strategy to construct the input representation.

[0015] For each input sample of the power load forecasting model to be explained, time series decomposition is performed; in constructing the conceptual alliance... The corresponding input representation is for members of the alliance. The concept uses the original input sample decomposition to obtain the corresponding trend component, seasonal component, and residual component; for those not belonging to the alliance The concept is to randomly select several samples from the training set, calculate the mean of their corresponding components, and use the mean to make replacements. The average value of the concept contribution of each component in all input samples in the dataset is calculated to obtain the overall contribution of each component in all samples, which is then output as the concept-level result.

[0016] Furthermore, the morphological interpretation results include: Construct a set of morphological prototypes, each prototype spanning a time period of [time span missing]. Then the set of morphological prototypes Represented as:

[0017] in They represent the first A prototype form; Multilayer perceptron is used to enhance the features of morphological prototype sequences:

[0018] in This represents the encoded morphological prototype. A sliding window approach is used to search for local structures in the input sequence that are similar to each prototype; let the input sequence be:

[0019] The subsequence of length L at time t is represented as: The subsequence is calculated by convolutional matching and compared with the first... k The similarity responses between morphological prototypes are used to construct a morphological activation matrix:

[0020] in Indicates the first k The prototype of each form in time position t The matching strength, where ; In obtaining the morphological activation matrix Then, feature aggregation is performed on the activation matrix along the prototype dimension to obtain the time series morphological representation:

[0021] This forms a sequence of morphological features. ,in They are respectively A time series format representation; The predicted output is obtained using a linear mapping. :

[0022] The contribution value of the target concept predicted by the morphological extraction network is used to obtain the morphological interpretation result.

[0023] Furthermore, the morphological prototype is learned by optimizing the objective function to minimize the distance between similar time series and the prototype, and maximize the distance between different classes. The steps include: The component with the largest contribution value in the conceptual explanation results is used as input data, and the calculated conceptual contribution value of the input sample of the power load prediction model to be explained is used as the optimization target output. An early shutdown strategy is adopted, and cross-validation is performed. The loss function includes task prediction loss, morphological matching loss, and prototype diversity loss. The joint optimization objective function achieves synergistic optimization of prediction accuracy and interpretation performance by weighting and balancing the task prediction loss, morphological matching loss, and prototype diversity loss.

[0024] in, , and These are the weights for task prediction loss, morphology matching loss, and prototype diversity loss, respectively. During training, the parameters of the encoding network, the morphological prototype vector P, and the parameters of the prediction projection layer are updated simultaneously using the gradient descent algorithm. Based on the training results, the best-performing optimal model is saved, and the optimal model parameters are recorded.

[0025] Furthermore, the obtained physical-level interpretation results include: Utilizing the central time Determine the set of absolute position indices of morphological activation within the original sample input window. ; index set Mapping back to the original physical feature space matrix By using index retrieval, physical variable subsequences synchronized with morphological time periods are extracted. It includes raw input features such as dry-bulb temperature, wet-bulb temperature, and humidity; Perform Min-Max normalization on the mapped physical variable subsequence to obtain the normalized load pattern fragment. : Normalized load pattern segment With physical variable subsequence By overlaying the images on a unified timeline and observing the correlation between their trends, we can analyze the physical driving factors of morphological changes and obtain physical-level explanations.

[0026] Furthermore, the multi-level interpretable analysis results of the generated power load forecasting model include: analyzing the changes in time characteristics, weather characteristics, and social dispatch characteristics in the original input data, and establishing the correspondence between morphological patterns and physical factors.

[0027] On the other hand, the present invention provides a hierarchical interpretable analysis system for power load forecasting models, comprising: The load history dataset construction module is used to collect historical power load data containing multi-dimensional features, perform multi-source data fusion and preprocessing, and construct a load history dataset. The prediction model training module is used to input the historical load dataset into the trained power load prediction model, obtain the load prediction sequence output by the power load prediction model, and define the feature space composed of the multi-dimensional features as the original feature space. The component decoupling module is used to decompose the historical power load data to obtain trend components, seasonal components, and dynamic residual components. The conceptual interpretation module is used to calculate the contribution of each component in the prediction output by perturbing or reconstructing the input of the trend component, seasonal component and dynamic residual component, and obtain the conceptual interpretation result. The morphological interpretation module is used to extract discriminative local morphological prototypes in the time series feature space corresponding to the load prediction sequence to obtain morphological interpretation results. The physical-level interpretation module is used to map the local morphological prototype to the original feature space, and based on the mapping result, to perform correlation analysis between the local morphological prototype and external physical variables to determine the dominant physical factors causing the local morphology and obtain the physical-level interpretation result. The results output module is used to integrate the conceptual-level interpretation results, morphological-level interpretation results and physical-level interpretation results to generate multi-level interpretable analysis results of the power load prediction model. The interpretation deployment module is used to construct a hierarchical interpretation process for load forecasting based on the results of the multi-level interpretable analysis, and to use the hierarchical interpretation process to perform real-time interpretation and analysis of online load forecasts.

[0028] Compared with the prior art, the present invention has the following beneficial effects: This invention proposes a cross-scale hierarchical interpretation method that can unify the modeling of macroscopic physical components and microscopic local morphology. Addressing the non-stationary nature of power load data, this invention employs a novel interpretation method that maintains the stability and consistency of interpretation results across different time distributions. By directly linking the interpretation results to the physical phenomena of power load, this invention imbues the prediction model's interpretation with clear physical semantics. Through direct correspondence with engineering phenomena such as load trend changes and peak-valley structures, this invention not only enhances the intuitiveness and understandability of the interpretation but also enables the interpretation results to directly provide practical basis for power dispatching and planning decisions, greatly improving the model's practicality and engineering application value. Attached Figure Description

[0029] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0030] Figure 1 This is a flowchart illustrating the implementation of an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the results in an embodiment of the present invention; Figure 3 This is a schematic diagram of the extraction results of typical load evolution patterns of trend components in an embodiment of the present invention; Figure 4 This is a schematic diagram of the cross-spatial time axis mapping result of the dataset in an embodiment of the present invention. Detailed Implementation

[0031] Through extensive research and practice, the inventors of this application have discovered that current load forecasting model interpretability methods face the following challenges: (1) The point-level interpretation granularity is low. Existing interpretation methods are usually limited to analyzing the point-level contribution of a single feature, failing to effectively combine macroscopic physical components and microscopic local morphology for unified modeling. Due to the lack of a hierarchical interpretation structure across scales, these methods struggle to capture the complex relationships between different levels in time series data, resulting in limitations on the accuracy and interpretability of the interpretation results. To gain a more comprehensive understanding of model behavior, multi-level fusion at the interpretation granularity is necessary to provide a richer analytical framework.

[0032] (2) Insufficient consistency in interpretation for load forecasting tasks. In regression tasks such as power load forecasting, due to the non-stationary nature of the data, existing interpretation methods based on perturbations or gradients are prone to fluctuations under different time distributions, which leads to the disruption of consistency between the interpretation results and the actual model behavior. Power load data has strong seasonality, trends, and random fluctuations. These factors make the attribution results generated by traditional interpretation methods unstable in different time periods, thus affecting the reliability and practicality of the interpretation.

[0033] (3) Insufficient physical semantic expression. Most current interpretation methods demonstrate the influence of features through numerical weights or importance scores. While these results can be quantified, they lack a direct connection with actual engineering phenomena of power load, such as load trend changes and peak-valley structure. Therefore, existing interpretation methods are difficult to provide intuitive and easily understandable basis for power dispatching and planning decisions, limiting their value and operability in practical applications. It is necessary to enhance the physical semantic expression to directly correspond the model output with the physical phenomena of the actual power system, thereby improving its application effect in real-world scenarios.

[0034] To achieve the above objectives, the main concept of this invention is as follows: First, a historical load dataset is constructed based on multidimensional data including historical load data, weather data, and socioeconomic data. Next, a target interpretive model is selected and trained on the constructed dataset, with optimal model parameters saved. Then, a trend-seasonal decomposition method based on locally weighted regression is used to extract the trend, seasonal, and residual components of the historical dataset. Next, three levels are constructed for the hierarchical interpretability method. In the conceptual interpretation module, a SHAP-based conceptual component attribution method is constructed to quantify the impact of different components on the prediction results. In the morphological interpretation module, a morphological extraction network is constructed to extract representative morphological prototypes from high-contribution components, characterizing key local change patterns. In the physical interpretation level, a time-axis mapping mechanism maps the morphological results back to the original load space, thereby establishing a correlation between the interpretive results and the actual trends of physical variable changes.

[0035] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0036] Example 1 like Figure 1As shown in the figure, this embodiment provides a hierarchical interpretable analysis method for power load forecasting models, which includes: Step 1: Collect historical power load data containing multi-dimensional features, and perform multi-source data fusion and preprocessing to construct a historical load dataset; Step 2: Input the historical load dataset into the trained power load prediction model, obtain the load prediction sequence output by the power load prediction model, and define the feature space composed of the multi-dimensional features as the original feature space. Step 3: Decompose the historical power load data to obtain trend components, seasonal components, and dynamic residual components; Step 4: By perturbing or reconstructing the inputs of the trend component, seasonal component, and dynamic residual component, calculate the contribution value of each component in the prediction output to obtain the conceptual interpretation results. Step 5: Extract discriminative local morphological prototypes from the time series feature space corresponding to the load prediction sequence to obtain morphological interpretation results; Step Six: Map the local morphological prototype to the original feature space, and perform correlation analysis between the local morphological prototype and external physical variables based on the mapping result to determine the dominant physical factors causing the local morphology and obtain the physical-level interpretation results. Step 7: Integrate the conceptual-level, morphological-level, and physical-level interpretation results to generate multi-level interpretable analysis results for the power load forecasting model; Step 8: Based on the multi-level interpretable analysis results, construct a hierarchical interpretation process for load forecasting, and use the hierarchical interpretation process to perform real-time interpretation and analysis of online load forecasting.

[0037] In step one of this embodiment, historical power load data with multi-dimensional features is first collected. This data includes time features (such as year, month, day, hour, and minute), weather features (such as temperature, humidity, and wind speed), and social dispatch features (such as electricity prices and holidays). By fusing data from different data sources, the original data is preprocessed to construct a historical load dataset. In one implementation, step one specifically includes: Step 11: In this embodiment, the dataset comes from user power load data in a certain region of Hebei Province, China, including eleven features: year, month, date, hour, minute, dry bulb temperature, dew point temperature, wet bulb temperature, humidity, electricity price, and power load. The time sampling interval is 15 minutes.

[0038] Steps one and two involve fusing load data from different data sources to construct a historical load dataset containing various feature dimensions. During data fusion, it is necessary to ensure the consistency of timestamps and to normalize data from different sources to ensure all features are on the same scale.

[0039] Step 13 involves dividing the preprocessed historical workload data into training, validation, and test sets. The training set is used for model training, the validation set is used for adjusting model parameters, and the test set is used for evaluating model performance.

[0040] In step two of this embodiment, the load dataset constructed in step one is trained using a deep learning model (such as the Transformer model) to model the load data, and the model is optimized using the training set. After training, the model's performance is evaluated using a validation set, and the trained load prediction model is finally saved. The model's performance is then tested by using the test set to predict future load trends and outputting the future load trend prediction results. In one implementation, step two specifically includes: Step 21: Select the Transformer deep learning model as the prediction model and train it using the training set. During training, mean squared error (MSE) is used as the loss function to optimize the model's predictive ability. Cross-validation and early stopping strategies are used to ensure the model's generalization ability on the validation set.

[0041] Step 22: Use the trained model to predict the test set and obtain the power load forecast results for future time periods. Test the model's prediction performance, verify its performance on unknown data, and save the optimal model obtained from training.

[0042] Step Two: Train the load dataset constructed in Step One using a deep learning model (such as the Transformer model) to model the load data, and optimize the model using the training set. After training, evaluate the model's performance using the validation set, and finally save the trained load prediction model. Test the model's performance by using the test set to predict future load trends and output the predicted future load trends. In step three of this embodiment, a load forecast sequence is generated based on the trained load forecasting model. Simultaneously, a seasonal trend decomposition method based on local weighted regression is used to decompose the input historical load data into trend components, seasonal components, and dynamic residual components. This step effectively extracts different features of the historical load data, which is helpful for subsequent detailed analysis of the forecast results. In one implementation, step three specifically includes: Step 31: The historical load data input to the model is decomposed using a seasonal trend decomposition method based on local weighted regression. The purpose of decomposition is to divide the original load series into trend components, seasonal components, and residual components, facilitating subsequent interpretation and analysis. During the decomposition process, a local weighted regression method is used to fit the trend and seasonal variations, and the residual components are calculated. Step 32 involves performing feature dimension pooling on the decomposed trend, seasonal, and residual components to reduce the dimensionality of the features, making them easier to input into the model for analysis. Feature dimension pooling helps reduce computational complexity and improves the efficiency of the subsequent interpretation module. In step four of this embodiment, the inputs of the trend component, seasonal component, and dynamic residual component are perturbed or reconstructed to calculate the contribution value of each component in the prediction output. Specifically, the marginal contribution of each concept (such as trend, seasonality, and residual) to the model prediction result is calculated, and precise mathematical formulas are used to quantitatively analyze the contribution weight of each component, ensuring stable and accurate contribution results. In one implementation, step four specifically includes: Step 41: Define the concept set and calculate the contribution value. Define the trend component, seasonal component, and dynamic residual component as three independent concepts. The size of the set is m=3; Step 42: For the given input samples of the power load forecasting model to be interpreted The load dataset constructed in step one is centrally selected, including historical load sequences and their corresponding time characteristics, meteorological characteristics, and social dispatch characteristics. After preprocessing, a fixed-length input data window is formed. A certain concept Contribution value The calculation formula is as follows:

[0043] in, It does not include All possible subsets of concepts It indicates the combination of concepts. The predicted output of the already trained power load forecasting model, i.e., incorporating concepts. The subsequent prediction results; In the combination of concepts The predicted output of the already trained power load forecasting model, i.e., without considering the concept. At that time, the prediction result is constructed by using a baseline replacement strategy to construct the input representation.

[0044] Step 43: For each sample to be explained, for the current concept... Construct a union state of two concepts. For a union state that does not contain any concepts... alliance , to the concept The corresponding time series components are replaced with the mean of the corresponding components from 100 randomly selected samples in the training set to simulate the model input when the concept is missing; for the added concept... The alliance that formed later If the concept is present, the time series component corresponding to that concept in the original input sample is retained to represent the model input when that concept exists. Then, the prediction models trained under both alliance states are input separately to obtain the corresponding prediction outputs. The difference between the model prediction results under the two states is then compared to calculate the concept's prediction. The marginal contribution to the model's prediction results. This method reduces evaluation bias caused by abrupt inputs and enhances the realism of the interpreted results by replacing the true component distributions from the training set with zero-value fillers.

[0045] Step 44: Calculate the mean of the conceptual contributions of the trend, seasonal, and residual components of all input samples in the dataset to obtain the overall contribution of each component in all samples.

[0046] In step five of this embodiment, discriminative local morphological prototypes are first extracted from the time series. Through training and inference with a morphological extraction network, the model learns K morphological prototypes and calculates the similarity between each subsequence and a prototype using convolutional matching, generating a morphological activation matrix. The matching strength between each subsequence and a prototype forms the elements of the matrix. Ultimately, through these morphological activation matrices, the model can accurately extract local time series morphologies highly correlated with power load forecast results.

[0047] In one implementation, step five specifically includes: Step 51: In the morphological interpretation module, the number of prototypes K and the prototype length L are first preset; the prototype parameters are randomly initialized using the Kaiming method; then, the backpropagation algorithm is used to jointly optimize the task prediction loss, morphological matching loss, and prototype diversity loss, continuously updating the prototype parameters so that each prototype gradually learns a typical local change pattern in the power load sequence, ultimately forming a morphological prototype vector P. Each morphological prototype corresponds to a typical local change pattern in the power load sequence. In this embodiment, the prototype vector is generated through Kaiming initialization and continuously optimized during training, gradually approximating the typical pattern in the real power load sequence. Prototype Set It can be represented as:

[0048] in This represents the k-th prototype form. In this embodiment, K is set to 6.

[0049] Step 52: Enhance the features of the morphological prototype sequence by encoding the prototype vector P using a multilayer perceptron (MLP) to obtain more discriminative morphological features, as shown in the following formula:

[0050] By mapping to a high-dimensional feature space, the distinguishability between different morphological prototypes is enhanced, and the interference of noise on the matching process is reduced.

[0051] Step 53: In the morphology matching stage of this embodiment, the model uses a sliding window approach to search for local subsequences similar to each morphological prototype in the historical load time series input of the sample to be explained. The similarity between each subsequence and the morphological prototype is calculated through convolutional matching to generate a morphological activation matrix A.

[0052]

[0053] in Let represent the matching strength of the k-th prototype at time position t, where .

[0054] The activation matrix A represents the matching strength of each prototype at each time position, describing the response distribution of different morphological patterns over the entire time series. In this embodiment, T is 96, representing a sample length of 96 time steps.

[0055] Step 54: Perform feature pooling aggregation on the morphological activation matrix A:

[0056] This forms a sequence of morphological features. .

[0057] Step 55: Subsequently, a linear mapping is used to obtain the predicted output, namely the target concept contribution value predicted by the morphological extraction network, which is then used to fit the true concept contribution value calculated in Step 4. :

[0058] The model uses the component with the largest contribution value in the conceptual explanation as input data and the calculated conceptual contribution value of each sample data as the target output, constructing training, validation, and test sets. The training set is used for training, and the validation set is used to evaluate the model's predictive performance. An early stopping strategy is employed, and cross-validation is used to further enhance the model's generalization ability. The loss function mainly consists of task prediction loss, morphological matching loss, and prototype diversity loss.

[0059] The task prediction loss is the core supervisory signal for interpreting the model, ensuring that the output accurately approximates the target observation. In this embodiment, the mean absolute error (MAE) is used as the objective function to enhance the model's robustness to outliers in non-stationary perturbations.

[0060] in, This represents the component contribution target attribution value of the i-th sample from the DDLFM model. By minimizing the error between the predicted output and the true observation, the morphological extraction network is trained, enabling it to learn key time-series morphological features relevant to the prediction task. The concept contribution value calculated based on the DDLFM model in step four is... This is used to characterize the degree of influence of different concepts on the model's prediction results, serving as a basis for measuring the importance of morphological patterns during the interpretation phase, but it is not included in the calculation of the prediction loss function. By combining prediction constraints and concept contribution analysis, it ensures that the extracted morphological patterns can effectively reflect the key influencing factors in the model's prediction process, thereby improving the fidelity of the interpretation results.

[0061] The morphological matching loss ensures that the learned morphological prototypes are not meaningless numerical noise, but rather derived from typical segments of the real load sequence. This loss constrains the prototypes to the maximum response subsequence in the input sequence using mean squared error (MSE). ,in, Indicates the first For each morphological prototype, the local time window corresponding to the maximum matching response in the input sequence is obtained. The distance between the morphological prototype and the corresponding maximum response subsequence is calculated using the mean squared error (MSE).

[0062] This prompts the prototype to approach the true form of time-series data, ensuring the physical authenticity of the morphological interpretation results.

[0063] The prototype diversity constraint is used to prevent multiple prototypes from learning the same repetitive patterns during model training. This term forces different prototypes to remain orthogonal or mutually exclusive in the feature space by maximizing the overall variance of the prototype set.

[0064] in, Represents the prototype diversity loss function; Indicates by A set of prototypes consisting of learnable morphological prototypes; Indicates the first A prototype of a form. ; Indicates the total number of morphological prototypes; It represents the variance of the prototype set in the feature space, and is used to measure the degree of difference between different prototypes; This represents a very small constant to prevent numerical computational instability. This mechanism ensures that the K prototypes can cover the widest possible range of data variations.

[0065] Ultimately, the joint optimization objective function achieves synergistic optimization of prediction accuracy and interpretation performance by weighting and balancing the aforementioned losses:

[0066] in, , and The learning level of the model for different objectives was controlled separately. Through parameter grid search experiments, in this embodiment, the final result was... Take it as 4. Take it as 6. Let it be 5.

[0067] During training, the model simultaneously updates the encoding network parameters, the morphological prototype vector P, and the projection layer parameters (which adjust the mapping between feature representations and predicted outputs) using the gradient descent algorithm, ultimately achieving end-to-end learning from abstract features to concrete forms. Based on the training results, the best-performing model is saved, and its optimal parameters are recorded.

[0068] In step six of this embodiment, the temporal position of the morphological prototype is first mapped back to the feature space of the original input data using a time-domain indexing and anchoring technique. By comparing the fluctuation trend of power load with changes in physical characteristics such as meteorological variables (e.g., temperature, humidity), the relationship between the prediction results and actual physical variables is revealed. Through time-domain alignment and normalization, the consistency of different features on the time axis is ensured, enabling effective comparison between morphological changes and fluctuations in physical variables, further improving the physical authenticity of the interpretation results.

[0069] In one implementation, step six specifically includes: Step 61: By calculating the position of the maximum response of the morphological prototype in the input time series, determine the time position corresponding to the morphological activation, and align the morphological activation with the physical variable subsequence within the original input window:

[0070] in, This represents the index of a time position in the input time series; This indicates the position of the maximum response of the morphological prototype in the input sequence, i.e., the center moment corresponding to the morphological activation; This indicates the length of the time window corresponding to the prototype.

[0071] Use the center time in the formula To determine the absolute location of morphological activation in the original sample.

[0072] Step 62: Map the time period corresponding to the morphological activation back to the original feature space using temporal indexing. Extract the corresponding physical variable subsequence (Dry bulb temperature, wet bulb temperature, humidity, electricity price); Step 63 involves performing Min-Max normalization on the extracted physical variable subsequences to ensure they are compared on a uniform scale. This step guarantees consistency in the fluctuation trends among different features.

[0073] Step 64: Normalize the load pattern fragments With physical variable sequence

[0074] By overlaying the images on a unified timeline and observing the correlation between their trends, we can help analyze the physical driving factors of morphological changes.

[0075] In step seven of this embodiment, the interpretation results from all three levels are ultimately merged to generate a complete hierarchical interpretable analysis result for the power load forecasting model. This analysis result not only provides power dispatchers with a clear basis for model decision-making but also helps them better understand the model's predictive performance in different time periods.

[0076] In one implementation, step seven specifically includes: In this embodiment, the final interpretation result is a synthesis of conceptual-level contribution results, morphological-level matching results, and physical-level correlation analysis results, providing comprehensive and interpretable analysis results for the power load forecasting model. These results offer more transparent decision support to power dispatchers and help them understand and optimize the behavior of the power load forecasting model.

[0077] This embodiment selects a multidimensional dataset of electricity load in Australia, containing nine features: date, hour, dry-bulb temperature, dew point temperature, wet-bulb temperature, humidity, electricity price, agricultural cycle, and electricity load, with a time interval of 15 minutes. The data was collected in a real-world environment, reflecting actual electricity consumption in daily life, enhancing the realism and practicality of the research; the dataset covers five years without any time gaps, facilitating analysis and modeling.

[0078] Figures 2 to 4 This demonstrates the multi-level interpretable analysis results of this embodiment on the Australia dataset. Figure 2This is a stacked plot showing the temporal evolution of the attribution weights of the components in the Australia dataset. It illustrates the changes in the attribution weights of the trend component, seasonal component, and dynamic component across different time samples, reflecting the dynamic evolution of the contribution of different conceptual components to the model's prediction results over time.

[0079] Figure 3 The results of extracting typical load evolution patterns from the trend components of the Australia dataset demonstrate the multiple typical shapelets learned by the model from the trend components. Different shapelets correspond to different load change patterns and are used to characterize the representative local evolution features of the time series identified by the model.

[0080] Figure 4 The cross-spatial-temporal mapping results for the Australia dataset demonstrate the alignment of extracted key load morphology segments with corresponding physical variables (including dry-bulb temperature, wet-bulb temperature, humidity, and load value) on a unified time axis. By comparing the consistency between load morphology changes and the trends of physical variable changes, the driving effect of different physical factors on load changes is analyzed, thus achieving a physical-level interpretation of the model prediction results.

[0081] In step eight of this embodiment, based on the above method, a hierarchical interpretation system for load forecasting is constructed. This system can process new power load data in real time and generate corresponding interpretable analysis reports based on the model's prediction results. Through its online analysis function, this system allows dispatchers to obtain detailed explanations of the model's predictions at any time, providing more transparent and reliable decision support.

[0082] Example 2 This embodiment provides a hierarchical interpretable analysis system for power load forecasting models, specifically including: The load history dataset construction module is used to collect historical power load data containing multi-dimensional features, perform multi-source data fusion and preprocessing, and construct a load history dataset. The prediction model training module is used to input the historical load dataset into the trained power load prediction model, obtain the load prediction sequence output by the power load prediction model, and define the feature space composed of the multi-dimensional features as the original feature space. The component decoupling module is used to decompose the historical power load data to obtain trend components, seasonal components, and dynamic residual components. The conceptual interpretation module is used to calculate the contribution of each component in the prediction output by perturbing or reconstructing the input of the trend component, seasonal component and dynamic residual component, and obtain the conceptual interpretation result. The morphological interpretation module is used to extract discriminative local morphological prototypes in the time series feature space corresponding to the load prediction sequence to obtain morphological interpretation results. The physical-level interpretation module is used to map the local morphological prototype to the original feature space, and based on the mapping result, to perform correlation analysis between the local morphological prototype and external physical variables to determine the dominant physical factors causing the local morphology and obtain the physical-level interpretation result. The results output module is used to integrate the conceptual-level interpretation results, morphological-level interpretation results and physical-level interpretation results to generate multi-level interpretable analysis results of the power load prediction model. The interpretation deployment module is used to construct a hierarchical interpretation process for load forecasting based on the results of the multi-level interpretable analysis, and to use the hierarchical interpretation process to perform real-time interpretation and analysis of online load forecasts.

[0083] It should be understood that any parts not described in detail in this specification belong to the prior art.

[0084] It should be understood that the above description of the preferred embodiments is quite detailed, but this should not be construed as limiting the scope of protection of this invention. It is neither necessary nor possible to exhaustively describe all possible implementations. Those skilled in the art, guided by this invention, can make substitutions or modifications without departing from the scope of the claims, all of which fall within the scope of protection of this invention. The scope of protection of this invention should be determined by the appended claims.

Claims

1. A hierarchical interpretable analysis method for power load forecasting models, characterized in that, Includes the following steps: Collect historical power load data containing multidimensional features, and perform multi-source data fusion and preprocessing to construct a historical load dataset; Input the historical load dataset into the trained power load prediction model, obtain the load prediction sequence output by the power load prediction model, and define the feature space composed of the multidimensional features as the original feature space. The historical power load data is decomposed to obtain trend components, seasonal components, and dynamic residual components; By perturbing or reconstructing the inputs of the trend component, seasonal component, and dynamic residual component, the contribution value of each component in the predicted output is calculated, and a conceptual interpretation result is obtained. Local morphological prototypes with discriminative capabilities are extracted from the time series feature space corresponding to the load prediction sequence to obtain morphological interpretation results; The local morphological prototype is mapped to the original feature space. Based on the mapping result, the local morphological prototype is correlated with external physical variables to determine the dominant physical factors that cause the local morphology and obtain physical-level interpretation results. By combining the conceptual, morphological, and physical interpretation results, multi-level interpretable analysis results of the power load prediction model are generated. Based on the results of the multi-level interpretable analysis, a hierarchical interpretation process for load forecasting is constructed, and the hierarchical interpretation process is used to perform real-time interpretation and analysis of online load forecasts.

2. The hierarchical interpretable analysis method for power load forecasting models according to claim 1, characterized in that, The multidimensional features include time features, weather features, and social scheduling features. The construction of the load dataset specifically includes: Collect historical power load data with multidimensional features, including time features such as year, month, day, hour, and minute; weather features such as temperature, humidity, and wind speed; and social dispatch features such as electricity price and holidays. Multi-source data fusion and preprocessing are performed on historical power load data from different data sources to construct a historical load dataset; The historical load dataset is divided into training set, validation set and test set.

3. The hierarchical interpretable analysis method for power load forecasting models according to claim 2, characterized in that: Obtaining the load forecast sequence output by the power load forecasting model includes: A deep load forecasting model is selected, trained based on the training set, with mean squared error as the loss function, and the effect is verified using a validation set. The trained model is then saved as the power load forecasting model. The test set is predicted based on the power load prediction model to obtain the load prediction sequence, and the prediction effect of the power load prediction model is tested.

4. The hierarchical interpretable analysis method for power load forecasting models according to claim 1, characterized in that, The decomposition of the historical power load data includes: The historical power load data input into the power load prediction model is decomposed using a seasonal trend decomposition method based on local weighted regression to obtain trend components, seasonal components, and residual components. Feature dimension pooling is performed on the trend component, seasonal component, and residual component to reduce the feature dimension.

5. The hierarchical interpretable analysis method for power load forecasting models according to claim 1, characterized in that, The contribution of each component to the predicted output is calculated, yielding conceptual-level explanations including: The trend component, seasonal component, and residual component are defined as three types of conceptual participants, forming a set. ,in These are the concepts of trend component, seasonal component, and residual component, respectively, with a set size of m=3; For a given input sample of the power load forecasting model to be interpreted , No. i One concept Contribution value The calculation formula is as follows: in, It does not include All possible subsets of concepts It indicates the combination of concepts. The predicted output of the already trained power load forecasting model, i.e., incorporating concepts. The subsequent prediction results; In the combination of concepts The predicted output of the already trained power load forecasting model, i.e., without considering the concept. At that time, the prediction result is constructed by using a baseline replacement strategy to construct the input representation; For each input sample of the power load forecasting model to be explained, time series decomposition is performed; in constructing the conceptual alliance... The corresponding input representation is for members of the alliance. The concept uses the original input sample decomposition to obtain the corresponding trend component, seasonal component, and residual component; for those not belonging to the alliance The concept is to randomly select several samples from the training set, calculate the mean of their corresponding components, and use the mean to make replacements. The average value of the concept contribution of each component in all input samples in the dataset is calculated to obtain the overall contribution of each component in all samples, which is then output as the concept-level result.

6. The hierarchical interpretable analysis method for power load forecasting models according to claim 1, characterized in that, The morphological interpretation results obtained include: Construct a set of morphological prototypes, each prototype spanning a time period of [time span missing]. Then the set of morphological prototypes Represented as: in They represent the first A prototype form; Multilayer perceptron is used to enhance the features of morphological prototype sequences: in This represents the encoded morphological prototype. A sliding window approach is used to search for local structures in the input sequence that are similar to each prototype; let the input sequence be: The subsequence of length L at time t is represented as: The subsequence is calculated by convolutional matching and compared with the first... k The similarity responses between morphological prototypes are used to construct a morphological activation matrix: in Indicates the first k The prototype of each form in time position t The matching strength, where ; In obtaining the morphological activation matrix Then, feature aggregation is performed on the activation matrix along the prototype dimension to obtain the time series morphological representation: This forms a sequence of morphological features. ,in They are respectively A time series format representation; The predicted output is obtained using a linear mapping. : The contribution value of the target concept predicted by the morphological extraction network is used to obtain the morphological interpretation result.

7. The hierarchical interpretable analysis method for power load forecasting models according to claim 6, characterized in that, The morphological prototype is learned by optimizing an objective function to minimize the distance between similar time series and the prototype, and maximize the distance between different classes. The steps include: The component with the largest contribution value in the conceptual explanation results is used as input data, and the calculated conceptual contribution value of the input sample of the power load prediction model to be explained is used as the optimization target output. An early shutdown strategy is adopted, and cross-validation is performed. The loss function includes task prediction loss, morphological matching loss, and prototype diversity loss. The joint optimization objective function achieves synergistic optimization of prediction accuracy and interpretation performance by weighting and balancing the task prediction loss, morphological matching loss, and prototype diversity loss. in, , and These are the weights for task prediction loss, morphology matching loss, and prototype diversity loss, respectively. During training, the parameters of the encoding network, the morphological prototype vector P, and the parameters of the prediction projection layer are updated simultaneously using the gradient descent algorithm. Based on the training results, the best-performing optimal model is saved, and the optimal model parameters are recorded.

8. The hierarchical interpretable analysis method for power load forecasting models according to claim 1, characterized in that, The physical-level interpretation results include: Utilizing the central time Determine the set of absolute position indices of morphological activation within the original sample input window. ; index set Mapping back to the original physical feature space matrix By using index retrieval, physical variable subsequences synchronized with morphological time periods are extracted. It includes raw input features such as dry-bulb temperature, wet-bulb temperature, and humidity; Perform Min-Max normalization on the mapped physical variable subsequence to obtain the normalized load pattern fragment. : Normalized load pattern segment With physical variable subsequence By overlaying the images on a unified timeline and observing the correlation between their trends, we can analyze the physical driving factors of morphological changes and obtain physical-level explanations.

9. A method for predicting rural power load based on a liquid neural network according to claim 1, characterized in that, The multi-level interpretable analysis results of the generated power load forecasting model include: analyzing the changes in time characteristics, weather characteristics and social dispatch characteristics in the original input data, and establishing the correspondence between morphological patterns and physical factors.

10. A hierarchical interpretable analysis system for power load forecasting models, characterized in that, Specifically, it includes: The load history dataset construction module is used to collect historical power load data containing multi-dimensional features, perform multi-source data fusion and preprocessing, and construct a load history dataset. The prediction model training module is used to input the historical load dataset into the trained power load prediction model, obtain the load prediction sequence output by the power load prediction model, and define the feature space composed of the multi-dimensional features as the original feature space. The component decoupling module is used to decompose the historical power load data to obtain trend components, seasonal components, and dynamic residual components. The conceptual interpretation module is used to calculate the contribution of each component in the prediction output by perturbing or reconstructing the input of the trend component, seasonal component and dynamic residual component, and obtain the conceptual interpretation result. The morphological interpretation module is used to extract discriminative local morphological prototypes in the time series feature space corresponding to the load prediction sequence to obtain morphological interpretation results. The physical-level interpretation module is used to map the local morphological prototype to the original feature space, and based on the mapping result, to perform correlation analysis between the local morphological prototype and external physical variables to determine the dominant physical factors causing the local morphology and obtain the physical-level interpretation result. The results output module is used to integrate the conceptual-level interpretation results, morphological-level interpretation results and physical-level interpretation results to generate multi-level interpretable analysis results of the power load prediction model. The interpretation deployment module is used to construct a hierarchical interpretation process for load forecasting based on the results of the multi-level interpretable analysis, and to use the hierarchical interpretation process to perform real-time interpretation and analysis of online load forecasting. The hierarchical interpretable analysis system for power load forecasting models is used to perform the steps in the hierarchical interpretable analysis method for power load forecasting models according to any one of claims 1-9.

Images