Load demand interval probability prediction method considering abnormal meteorological conditions

Abstract

The invention discloses a load demand interval probability prediction method considering abnormal meteorological conditions, and relates to the field of power load prediction. Obtaining multi-dimensional data and carrying out data fusion processing to obtain a multi-dimensional feature set; performing dimension reduction on the multi-dimensional feature set by using a principal component analysis method to extract all meteorological factors influencing load demand changes in historical meteorological data under an abnormal meteorological condition; analyzing the combined features of all meteorological factors to obtain all extreme meteorological scenes under abnormal meteorological conditions; all the extreme meteorological scenes are clustered to obtain different types of extreme meteorological scenes, and key factors influencing load demand changes under the different types of extreme meteorological scenes are identified; constructing time sequence models corresponding to different types of extreme meteorological scenes based on the key factors; and combining quantile regression with a time sequence model to construct an interval load prediction model. According to the invention, the prediction precision and stability of the load demand under the abnormal condition are improved.

Description

Technical Field

[0001] The present invention relates to the field of electric load forecasting, and more particularly, to a method for interval probability forecasting of load demand considering abnormal meteorological conditions. Background Art

[0002] Most of the existing load demand forecasting methods are based on deterministic models, which are difficult to effectively characterize the uncertainty of load demand under abnormal conditions, resulting in large forecasting errors. At the same time, the existing methods often rely on historical data for load forecasting, but in abnormal situations such as extreme weather and sudden accidents, the applicability of historical data is limited and cannot accurately reflect the changes in load demand. Summary of the Invention

[0003] In view of the above problems, the present invention proposes a method for interval probability forecasting of load demand considering abnormal meteorological conditions to overcome the deficiencies of the prior art.

[0004] An embodiment of the present invention provides a method for interval probability forecasting of load demand considering abnormal meteorological conditions, and the load demand interval probability forecasting method includes: Obtain multi-dimensional data and perform data fusion processing to obtain a multi-dimensional feature set, where the multi-dimensional data includes: historical meteorological data; Use the principal component analysis method to reduce the dimension of the multi-dimensional feature set, and extract all meteorological factors that affect the change of load demand under abnormal meteorological conditions in the historical meteorological data; Analyze the combined features of all meteorological factors to obtain all extreme meteorological scenarios under abnormal meteorological conditions; Cluster all the extreme meteorological scenarios to obtain different types of extreme meteorological scenarios, and identify the key factors that affect the change of load demand under different types of extreme meteorological scenarios; Based on the key factors, construct a time series model corresponding to different types of extreme meteorological scenarios, and this time series model is used to predict the average value of load demand at any future time point; Adopt quantile regression combined with the time series model to construct an interval load forecasting model, and this interval load forecasting model is used to predict the interval probability of load demand at any future time point.

[0005] Optionally, the multi-dimensional data further includes: historical load data, historical power grid operation status data; the historical meteorological data includes: abnormal meteorological data and normal meteorological data; Obtain multi-dimensional data and perform data fusion processing to obtain a multi-dimensional feature set, including: Adopt wavelet transform denoising and clustering anomaly detection methods to identify and eliminate outliers in the historical meteorological data, the historical load data, and the historical power grid operation status data, and obtain correct data; Use time synchronization processing and data methods to fuse the correct data, match the historical load data and the historical power grid operation status data with the corresponding abnormal meteorological data or normal meteorological data, and obtain the correlation between the abnormal meteorological data or the normal meteorological data and the load demand change at each time point; Label the correlation between the abnormal meteorological data and the load demand change as data under abnormal meteorological conditions; Label the correlation between the normal meteorological data and the load demand change as data under normal meteorological conditions.

[0006] Optionally, use the principal component analysis method to reduce the dimension of the multi-dimensional feature set, and extract all meteorological factors in the historical meteorological data that affect the load demand change under abnormal meteorological conditions, including: Convert the multi-dimensional feature set to a new low-dimensional feature space through the principal component analysis method to obtain a low-dimensional feature set, so that the features corresponding to each data after dimension reduction are orthogonal to each other and do not contain duplicate information; Based on the low-dimensional feature set, calculate the non-linear correlation between the load demand change and the historical meteorological data to obtain all the meteorological factors.

[0007] Optionally, based on the low-dimensional feature set, calculate the non-linear correlation between the load demand change and the historical meteorological data to obtain all the meteorological factors, including: Adopt the combination of correlation coefficient and non-linear metric to calculate the non-linear correlation between the load demand change and the historical meteorological data; Use Fourier transform to convert the non-linear correlation between the load demand change and the historical meteorological data from the time domain to the frequency domain, extract the periodic characteristics of each data, and identify all meteorological factors that have a non-linear impact on the load demand change under the abnormal meteorological conditions; Among them, the periodic characteristics are used as the step size when constructing the time series model.

[0008] Optionally, analyze the combined characteristics of all meteorological factors to obtain all extreme meteorological scenarios under abnormal meteorological conditions, including: Construct combinations of all meteorological factors to obtain combined factors; Analyze and capture the interaction between the combined factors to obtain all the extreme meteorological scenarios.

[0009] Optionally, cluster all the extreme meteorological scenarios to obtain different types of extreme meteorological scenarios, including: Cluster all the extreme meteorological scenarios by using one or more of k-means, k-means++, k-medoids, and k-median to obtain the different types of extreme meteorological scenarios; For each type of extreme meteorological scenario, identify the key factors that affect the change in the load demand.

[0010] Optionally, the different types of extreme meteorological scenarios include: extremely hot meteorological scenarios and extremely cold meteorological scenarios; The combined factors corresponding to the extremely hot meteorological scenario are:

[0011] The combined factors corresponding to the extremely cold meteorological scenario are:

[0012] In the above two formulas, represents the heat index, represents the cold index, represents the Fahrenheit temperature, represents the wind speed, represents the relative humidity, represents a constant; when the heat index exceeds the heat threshold, it is determined as the extremely hot meteorological scenario; when the cold index exceeds the cold threshold, it is determined as the extremely cold meteorological scenario.

[0013] Optionally, construct time series models corresponding to the different types of extreme meteorological scenarios based on the key factors, including: For each type of extreme meteorological scenario, input its corresponding key factors, the historical power load data extracted from the historical load data, and the static feature data extracted from the historical power grid operation status data into the time series model according to the step size, and use a long short-term memory network combined with an attention mechanism to construct a time series model corresponding to each type of extreme meteorological scenario; Among them, the attention mechanism is a multi-head parallel attention mechanism; the steps of constructing the time series model by combining the long short-term memory network and the attention mechanism include: Select a vector sequence input to the attention mechanism from the time series feature data at the step size, and each vector in the vector sequence is a d-dimensional vector containing the key factor, its corresponding historical power load data, and static feature data; Generate corresponding query vectors, key vectors, and value vectors for each vector; In the vector sequence of each head in the multi-head parallel attention mechanism, the query vector and key vector of each vector are respectively operated to obtain their respective attention scores; The attention score and value vector of each vector are converted into a weight distribution through Softmax to obtain the output of each head; The outputs of all heads are concatenated to form a high-dimensional output vector; Taking the high-dimensional output vector as the input of the long short-term memory network to form the time series model.

[0014] Optionally, a quantile regression is combined with the time series model to construct an interval load forecasting model, including: At the output level of the time series model corresponding to each type of extreme meteorological scenario, a quantile regression method is introduced to expand the average value output at its output level into an interval probability prediction, and an interval load forecasting model corresponding to each type of extreme meteorological scenario is constructed.

[0015] Optionally, after constructing the interval load forecasting model, it further includes: Optimizing the interval load forecasting model by using the historical meteorological data and historical load data.

[0016] The load demand interval probability forecasting method considering abnormal meteorological conditions proposed by the present invention creatively deals with abnormal meteorological conditions such as extreme weather, uses multi-dimensional data, and constructs a comprehensive feature extraction mechanism through data fusion technology. Introducing a multi-modal fusion technology to improve data utilization efficiency, and optimizing data features through a dimensionality reduction method to improve prediction accuracy. Quantifying the impact of the combined characteristics of meteorological factors on extreme meteorology and applying it to load demand forecasting is an important improvement to the existing forecasting methods.

[0017] By introducing the attention mechanism into the LSTM (long short-term memory network) and using quantile regression as the loss function of the prediction model, the ability of the model to capture long-time series dependence relationships is enhanced, the accuracy of load time series feature extraction is improved, and the complex time series dynamics and non-linear features of load data are fully captured. Using quantile regression as the loss function to construct an interval prediction model for load demand, providing a confidence interval for load demand, and effectively quantifying the uncertainty of the load. Solving the deficiencies of the existing methods, improving the prediction accuracy and stability of load demand under abnormal conditions, and providing more accurate decision-making support for power system scheduling.

[0018] The method proposed by the present invention is applicable to load forecasting under abnormal conditions, such as special scenarios like extreme weather, power grid failures, and new energy fluctuations. It can solve the long-term dependence relationship of traditional load methods on load data. By performing interval forecasting on the load, it can effectively characterize the non-linear dependence relationship of the load under abnormal conditions, and can provide a more reliable decision-making basis for power dispatching optimization, power market transactions, and new energy consumption, with broad application prospects and high practicality. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] By reading the following detailed description of the preferred embodiments, various other advantages and benefits will become clear to those of ordinary skill in the art. The drawings are only for the purpose of showing the preferred embodiments and are not considered as a limitation of the present invention. Moreover, throughout the drawings, the same reference numerals are used to represent the same components. In the drawings: Figure 1 is an architecture diagram of a bus-type valve group intelligent management system proposed by an embodiment of the present application; Figure 2 is an overview diagram of the process of combining long short-term memory networks with an attention mechanism in an embodiment of the present application; Figure 3 is an operation flow chart of introducing an attention mechanism in an embodiment of the present application; Figure 4 is an architecture diagram of an interval load forecasting model in an embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0020] To make the above objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below with reference to the drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, and are only a part of the embodiments of the present invention, rather than all of the embodiments, and are not used to limit the present invention.

[0021] Power load forecasting is an important basis for power system dispatching, optimal operation, and power market transactions. However, under abnormal conditions such as extreme weather, equipment failures, and sudden accidents, the load demand in the power receiving area fluctuates greatly. Traditional load forecasting methods cannot accurately reflect the volatility and uncertainty of the load, resulting in greater risks in power system dispatching and decision-making. In addition, with the rapid development of a high proportion of new energy, especially the grid-connected power generation of renewable energy such as wind power and photovoltaic power, the uncertainty, volatility, and reverse peak shaving characteristics of the output of wind and solar power generation in the system make the load balance of the power grid more complex.

[0022] The inventors found that many traditional models are trained based on the relationship between normal meteorology and load in previous years. When encountering extreme weather beyond the scope of experience, these models often fail to accurately depict the response of the load to abnormal climate and tend to underestimate (or overestimate) the load peak. Under meteorological conditions, the error of load forecasting will increase significantly. Traditional methods lack special treatment for extreme events. When rare weather conditions occur, the actual load often deviates greatly from the model prediction value. In this case, the "long-tail" characteristic of the model residual distribution becomes obvious, that is, the probability of extreme errors increases, resulting in unreliable prediction results.

[0023] The further penetration of new energy has led to greater dispatching difficulties for power systems, and traditional load forecasting methods have poor adaptability to a high proportion of new energy. Under normal circumstances, such point forecasts may be sufficient for reference. However, in extreme weather scenarios, especially in the context of a high proportion of new energy, only giving a single value will pose risks to power dispatching - operators cannot intuitively understand the possible range of prediction deviations from it, nor can they evaluate the probability of abnormal deviations. How to accurately predict load demand and provide a prediction interval has become an important issue in the optimization of power system dispatching.

[0024] The inventors further studied and found that there are various problems in the currently adopted load demand forecasting methods. For example, probability forecasting based on traditional time series models, such as autoregression and autoregressive moving average, mostly relies on historical load data for modeling and is usually applicable to situations where load changes are relatively regular. However, under abnormal conditions, its prediction accuracy often drops significantly.

[0025] Machine learning methods, such as random forest, gradient boosting decision tree, deep neural network, etc., all rely on a large amount of historical data and have poor generalization ability under abnormal conditions, making it difficult to accurately predict the load demand interval. Traditional point forecasting methods usually assume that load changes have a certain degree of stationarity or predictability, while actual load demand is affected by factors such as weather, holidays, and sudden accidents and fluctuates greatly. Under abnormal conditions (such as extreme weather, sudden accidents, equipment failures, etc.), the load demand often fluctuates greatly. Traditional point forecasting methods can only provide a deterministic value and cannot give the confidence interval of the prediction. Moreover, the point forecasting method only outputs the most likely prediction value, and it cannot cover abnormal load changes in extreme cases.

[0026] In view of the above problems, the inventors creatively proposed a method for probabilistic prediction of load demand interval considering abnormal meteorological conditions of the present invention. The technical solutions of this application will be explained and described in detail below.

[0027] A method for probabilistic prediction of load demand interval considering abnormal meteorological conditions proposed by the present invention refers to Figure 1 the flowchart shown, which includes: Step 101: Obtain multi-dimensional data and perform data fusion processing to obtain a multi-dimensional feature set. The multi-dimensional data includes: historical meteorological data.

[0028] The method proposed in the present invention aims at the above-mentioned problems and proposes an interval load forecasting method based on meteorological factors and historical load data, aiming to solve the problem that the existing power load forecasting methods cannot accurately reflect the load volatility and uncertainty under extreme meteorological conditions, especially in extreme high temperature, cold snap and other weather conditions. Most traditional load forecasting methods adopt point forecasting mode, which cannot provide the interval range of load demand and lack the ability to cope with uncertainty and extreme weather events.

[0029] To achieve the above object and implement the method proposed in the present invention, it is first necessary to obtain multi-dimensional data and perform data fusion processing to obtain a multi-dimensional feature set. The multi-dimensional data includes data in multiple dimensions such as historical meteorological data, historical load data, and historical power grid operation status data. These multi-dimensional data form a multi-dimensional feature set. The amount of data contained in this multi-dimensional feature set is large, and the data is independent and has no correlation. And there must be incorrect data such as outliers, and it also contains data that does not involve relevant factors such as meteorological factors and loads. Therefore, these data need to be screened.

[0030] A relatively optimal method for obtaining multi-dimensional data and performing data fusion processing to obtain a multi-dimensional feature set includes: Adopt wavelet transform denoising and clustering anomaly detection methods to identify and remove outliers in historical meteorological data, historical load data, and historical power grid operation status data to obtain correct data; then use time synchronization processing and data methods to fuse the correct data, and match the historical load data, historical power grid operation status data with corresponding abnormal or normal meteorological data to obtain the correlation between abnormal or normal meteorological data and load demand changes at each time point.

[0031] The above multi-dimensional feature set includes but is not limited to: meteorological factors such as temperature, humidity, wind speed, precipitation, and air pressure; load historical data includes but is not limited to: daily load, monthly load, weekly load, etc.; power grid operation status data includes but is not limited to: generation load, transmission line status, load distribution, etc.

[0032] Extreme meteorological conditions will significantly affect load fluctuations. For example, heatwaves lead to a sharp increase in cooling loads and load peaks; cold snaps increase heating demands and cause sudden increases in power loads, etc. Use time synchronization processing and data fusion to match the power grid load data with the corresponding extreme meteorological data, so that the load data at each time point is associated with the corresponding meteorological characteristics. At the same time, adopt wavelet transform denoising and clustering anomaly detection methods to identify and remove outliers to ensure data quality.

[0033] After completing the above content, the correlation between the tags of abnormal meteorological data and the change in load demand is the data under abnormal meteorological conditions; the correlation between the tags of normal meteorological data and the change in load demand is the data under normal meteorological conditions. In this way, the data involved in these two aspects of abnormal and normal are distinguished, laying a foundation for the subsequent implementation of the method proposed in the present invention.

[0034] Step 102: Use the principal component analysis method to reduce the dimension of the multi-dimensional feature set, and extract all meteorological factors that affect the change in load demand under abnormal meteorological conditions from the historical meteorological data.

[0035] Since the historical load data with a huge amount of data and various related features (such as temperature, humidity, wind speed, etc.) are often highly redundant or correlated, directly using these data may lead to too high complexity of the subsequent model and introduce noise. The dimensionality reduction technology can compress the feature space while retaining the key information and reduce the data dimension.

[0036] Therefore, after step 101 processes the multi-dimensional feature set, in order to reduce data redundancy and improve the training efficiency of the subsequent model, it is necessary to use the principal component analysis method to reduce the dimension of the multi-dimensional feature set, and extract all meteorological factors that affect the change in load demand under abnormal meteorological conditions from the historical meteorological data. Of course, it can be understood that the same method can also be used to obtain all meteorological factors that affect the change in load demand under normal meteorological conditions, or other known technologies can also be used.

[0037] A preferred method for extracting all meteorological factors that affect the change in load demand under abnormal meteorological conditions from the historical meteorological data includes: Convert the multi-dimensional feature set to a new low-dimensional feature space through the principal component analysis method to obtain a low-dimensional feature set, so that the features corresponding to each data after dimensionality reduction are orthogonal to each other and do not contain duplicate information; then, based on the low-dimensional feature set, calculate the non-linear correlation between the change in load demand and the historical meteorological data to obtain all meteorological factors. Specifically: The non-linear correlation between the change in load demand and the historical meteorological data can be calculated by combining the correlation coefficient with the non-linear metric; then, use the Fourier transform to convert the non-linear correlation between the change in load demand and the historical meteorological data from the time domain to the frequency domain, extract the periodic features of each data, and identify all meteorological factors that have a non-linear impact on the change in load demand under abnormal meteorological conditions; among them, the periodic feature is used as the step size when constructing the time series model.

[0038] The principle of the above-mentioned principal component analysis method is summarized as follows: Cv i =λi v i (3) Y i = X i V k (4) In the above formula, μ represents the mean of each feature, S 2 represents the standard deviation, X i is the i-th sample value, λ i is the i-th eigenvalue, v i represents the corresponding eigenvector, represents the data after dimensionality reduction.

[0039] Equation (1) standardizes each feature to eliminate the influence of dimension. Equation (2) calculates the covariance matrix of the standardized data. Equation (3) calculates the eigenvalues and eigenvectors of the covariance C matrix. By sorting the calculated eigenvalues, the first k are selected as eigenvectors. Equation (4) forms a transformation matrix by arranging the selected k eigenvectors in columns, and then projects the original data onto the new dimensional space.

[0040] In feature engineering, calculating the non-linear correlation between load and meteorological factors helps to reveal relationships that are difficult to discover through ordinary linear analysis. In extreme meteorological scenarios, the correlation coefficient combined with non-linear metrics is used to evaluate the degree of association between multi-dimensional meteorological features and load, so as to identify the non-linear impact of extreme weather conditions on load fluctuations.

[0041] (5) In the above formula, , respectively represent the , th variable sample values, , are the means of variables , respectively, and the numerator , is the covariance, and the denominator is the product of the standard deviations of , .

[0042] Since the electric load has significant periodic patterns, in extreme meteorological scenarios, its periodic characteristics may change abnormally. The Fourier transform is used to convert time series data from the time domain to the frequency domain, extract the periodic characteristics of the data, and identify the impact of extreme weather or special events on load fluctuations.

[0043] (6) In the above formula, represents the th frequency component of the frequency-domain signal, represents the th sampling point of the time-domain signal, represents the length of the signal, represents the imaginary unit, represents the rotation factor, which is used for spectrum calculation.

[0044] In the above way, the non-linear correlation between the load demand change and the historical meteorological data can be calculated to obtain all meteorological factors.

[0045] Step 103: Analyze the combined characteristics of all meteorological factors to obtain all extreme meteorological scenarios under abnormal meteorological conditions.

[0046] After obtaining all meteorological factors, it is also necessary to analyze the combined characteristics of all meteorological factors to obtain all extreme meteorological scenarios under abnormal meteorological conditions. By using the correlation between meteorological factors, the extreme meteorological scenarios that may cause large fluctuations in load demand are identified, providing a theoretical basis for load forecasting under extreme weather conditions.

[0047] A better method for analyzing the combined characteristics of all meteorological factors to obtain all extreme meteorological scenarios under abnormal meteorological conditions includes: Construct combinations of all meteorological factors to obtain combined factors; analyze and capture the interaction between these combined factors to obtain all extreme meteorological scenarios.

[0048] Preferably, different types of extreme meteorological scenarios include: extremely hot meteorological scenarios and extremely cold meteorological scenarios.

[0049] The combined factor corresponding to the extremely hot meteorological scenario is:

[0050] The combined factor corresponding to the extremely cold meteorological scenario is:

[0051] In the above two equations, represents the heat index, represents the cold index, represents the Fahrenheit temperature, represents the wind speed, represents the relative humidity, represents a constant; when the heat index exceeds the heat threshold, it is determined as an extremely hot meteorological scenario; when the cold index exceeds the cold threshold, it is determined as an extremely cold meteorological scenario.

[0052] Step 104: Cluster all extreme meteorological scenarios to obtain different types of extreme meteorological scenarios, and identify the key factors that affect the change of load demand under different types of extreme meteorological scenarios.

[0053] After obtaining all extreme meteorological scenarios, it is also necessary to cluster all extreme meteorological scenarios to obtain different types of extreme meteorological scenarios, and identify the key factors that affect the change of load demand under different types of extreme meteorological scenarios. As can be seen from the foregoing: for different types of extreme meteorological scenarios, the corresponding key factors that affect the change of load demand may be different. Therefore, it is necessary to cluster all extreme meteorological scenarios to obtain different types of extreme meteorological scenarios, and identify the key factors that affect the change of load demand under different types of extreme meteorological scenarios.

[0054] Preferably, one or more of k-means, k-means++, k-medoids, and k-median can be used to cluster all extreme meteorological scenarios to obtain different types of extreme meteorological scenarios.

[0055] Step 105: Build time series models corresponding to different types of extreme meteorological scenarios based on the key factors. The time series models are used to predict the average value of load demand at any future time point.

[0056] After the above steps 101 to 104 are completed, build time series models corresponding to different types of extreme meteorological scenarios based on the key factors, that is: for each type of extreme meteorological scenario, a corresponding time series model needs to be built, and the time series model is used to predict the average value of load demand at any future time point.

[0057] A preferred method for building time series models corresponding to different types of extreme meteorological scenarios includes: For each type of extreme meteorological scenario, input its corresponding key factors, historical power load data extracted from historical load data, and static feature data extracted from historical power grid operation state data into the time series model according to a step size, and use a long short-term memory network combined with an attention mechanism to build a time series model corresponding to each type of extreme meteorological scenario.

[0058] Build a time series prediction model with a long short-term memory network as the core, and use its gating structure to capture the long-term dependence relationship of the load sequence. However, when building a model with a long short-term memory network, due to information attenuation during the extraction of long sequence information, as well as problems such as severe fluctuations and abnormal patterns of load under extreme meteorological scenarios, the time series model may not be able to fully utilize the effective historical information of long time periods to predict future loads. To solve this problem, the inventor creatively proposes to introduce an attention mechanism on the basis of the long short-term memory network.

[0059] The attention mechanism enables the time series model to automatically focus on the more relevant critical moments or features in the historical sequence during prediction, thereby enhancing the sensitivity to abnormal patterns. This structure effectively alleviates the information attenuation during the extraction of long sequence information, as well as reduces problems such as drastic fluctuations in load and abnormal patterns under extreme meteorological scenarios, enabling the model to make more full use of the effective historical information in the long time period to predict future load.

[0060] The attention mechanism is a multi-head parallel attention mechanism; a preferred step for constructing a time series model by combining a long short-term memory network and an attention mechanism includes: Select a vector sequence input to the attention mechanism from the time series feature data at a step size. Each vector in this vector sequence is a d-dimensional vector containing key factors, their corresponding historical power load data, and static feature data. Generate corresponding query vectors, key vectors, and value vectors for each vector; calculate the respective attention scores based on the query vectors and key vectors of each vector in the vector sequence of each head in the multi-head parallel attention mechanism.

[0061] Convert the respective attention scores and value vectors of each vector into a weight distribution through Softmax to obtain the output of each head; splice the outputs of each group of all heads to form a high-dimensional output vector; use the high-dimensional output vector as the input of the long short-term memory network to form a time series model.

[0062] To better understand the above process of combining a long short-term memory network and an attention mechanism, refer to Figure 2 the overview diagram of the process of combining a long short-term memory network and an attention mechanism shown in Figure 3 the operation flowchart of introducing the attention mechanism shown in Using time series feature data (which is comprehensive data of relevant key factors and historical power load data, such as historical power load values and meteorological weather data , spectral features and so on) and static feature data

[0063] During the process of constructing a time series model, using the above as the input feature set ( Figure 2 in ), when integrating temporal features and static features, it adopts the Attention mechanism to highlight the importance of key time steps or key features, and then inputs the extracted global context information into a multi-layer LSTM network for deep temporal modeling.

[0064] Static features: represented by , include features that do not change frequently over time or change only over a long period, such as: regional information, device features, climate zone classification, etc. Temporal features: represented by and . Here represents the spectral feature, while is the main target sequence or associated basic temporal feature (such as the load value of the previous few days, the change of weather features over time, etc.). Use to determine the input step size, that is, how much data to select from to input into the Attention mechanism. For example: F 1 = 10, then select from the range of x 1 to x 10 and input it into the Attention mechanism. F 2 = 100, then select from the range of x 1 to x 100 and input it into the Attention mechanism.

[0065] The input sequence is , and each is a d-dimensional vector (n-dimensional time steps, determined by , and each time step contains d-dimensional information, that is, the feature information screened by the previous features, including power load / temperature / wind speed / humidity, etc.). Generate corresponding query (Query) vectors, key (Key) vectors, and value (Value) vectors for each vector. There is a d-dimensional learnable parameter matrix, W Q , W K , W V , and for each head, it is necessary to map the input vector to the query vector q1h, and then map out their respective keys and values.

[0066] For each in the sequence, , , perform three groups of independent linear transformations, usually represented by X , and then the calculation formula is unified as:

[0067]

[0068]

[0069] For a single head, that is, each head: The query vector and key vector of each vector are respectively operated to obtain their respective attention scores (a 1，i ). To calculate the matching degree of q i with the entire sequence , and then divide by the scaling factor (dk is the mapping dimension of the query and the key) to obtain the attention score at the i-th position , and then convert it to a weight distribution through Softmax , that is: .

[0070] The previous step only described the output process of one attention head e 1. In multi-head attention, there are a total of h parallel heads, and each head will obtain an independent output: e 2, e 3, …… e n , e s . Concatenate these outputs to form a higher-dimensional vector, and then input it into the LSTM layer for time prediction to form a time series model

[0071] For LSTM, each single layer has: Forget gate: used to determine how much cell state information from the previous moment to retain:

[0072] Input gate: determines which information needs to be input into the cell state at the current moment:

[0073]

[0074] Output gate: outputs the hidden state:

[0075]

[0076] Update the cell state: Combine the outputs of the input gate and the forget gate to update the cell state at the current moment:

[0077] In the above formula, represents the activation function represents the weight matrix of the forget gate, represents the bias of the forget gate, represents the output of the input gate, represents the candidate cell state, represents the forget gate output, represents a constant, represents the output of the output gate, represents the non - linear activation of the current cell state, represents a constant.

[0078] After passing and multi - layer combination, stack multiple LSTM layers together to capture deeper temporal patterns in the data. After stacking multiple LSTM layers, the hidden state sequence of the last layer finally obtained can be used for the prediction output of subsequent tasks.

[0079] Step 106: Combine quantile regression with the time - series model to construct an interval load prediction model, which is used to predict the interval probability of the load demand at any future time point.

[0080] After obtaining the time - series model in Step 105, it can only predict the average value of the load demand at any future time point. In extreme scenarios, point prediction alone is difficult to meet the decision - making needs because point prediction cannot reflect the uncertainty of future loads. It cannot predict the upper and lower limit values of the load demand at any future time point, that is, it cannot predict the interval probability of the load demand at any future time point. To achieve this goal, by training models for different quantiles, the load prediction is extended to probability prediction, providing an interval range rather than a single value for each moment. That is: combine quantile regression with the time - series model to construct an interval load prediction model, which is used to predict the interval probability of the load demand at any future time point.

[0081] Refer to Figure 4 the architecture diagram of the interval load prediction model shown: At the output level of the time - series model corresponding to each type of extreme meteorological scenario, introduce the quantile regression method ( Figure 4 denoted by Dropout in

[0082] to expand the average value output at its output level to interval probability prediction, and construct the interval load prediction model corresponding to each type of extreme meteorological scenario. Introduce quantile regression as the loss function, and the formula is as follows: In the formula, represents the loss function, which is used to calculate the loss value of the th data point at time under the quantile is the predicted value, representing the th data point at time , and the prediction result based on the quantile . is the actual value, representing the th data point at time . represents the target quantile. The final output interval load prediction Y min , Y max = f ( X , S ).

[0083] Through the above method, an interval load prediction model can be finally obtained, which can predict the interval probability of the load demand at any future time point. In order to obtain a more accurate prediction result, the interval load prediction model constructed can also be continuously optimized by using historical meteorological data and historical load data, so as to accurately predict the interval probability of the load demand at any future time point.

[0084] The load demand interval probability prediction method considering abnormal meteorological conditions proposed by the present invention creatively considers abnormal meteorological conditions such as extreme weather, uses multi-dimensional data, and constructs a comprehensive feature extraction mechanism through data fusion technology. The multi-modal fusion technology is introduced to improve the data utilization efficiency, and the data features are optimized by the dimensionality reduction method to improve the prediction accuracy. The combined characteristics of meteorological factors are used to quantify the impact of extreme meteorology and applied to the load demand prediction, which is an important improvement to the existing prediction methods.

[0085] By introducing the attention mechanism into LSTM (Long Short-Term Memory Network) and using quantile regression as the loss function of the prediction model, the ability of the model to capture long-term time series dependence is enhanced, the accuracy of load time series feature extraction is improved, and the complex time series dynamics and non-linear features of load data are fully captured. Using quantile regression as the loss function, an interval prediction model of load demand is constructed to provide the confidence interval of load demand and effectively quantify the uncertainty of load. Solve the deficiencies of the existing methods, improve the prediction accuracy and stability of load demand under abnormal conditions, and provide more accurate decision-making support for power system scheduling.

[0086] The method proposed by the present invention is applicable to load forecasting under abnormal conditions, such as special scenarios like extreme weather, power grid failures, and new energy fluctuations. It can solve the long-term dependence relationship of traditional load methods on load data. By performing interval forecasting on the load, it can effectively characterize the non-linear dependence relationship of the load under abnormal conditions, and can provide a more reliable decision-making basis for power dispatching optimization, power market trading, and new energy consumption, with broad application prospects and high practicality.

[0087] Although the preferred embodiments of the embodiments of the present invention have been described, those skilled in the art can make additional changes and modifications to these embodiments once they know the basic creative concept. Therefore, the appended claims are intended to be construed as including the preferred embodiments and all changes and modifications falling within the scope of the embodiments of the present invention.

[0088] Finally, it should also be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or terminal device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or terminal device. Without further limitation, an element defined by the statement "comprising one..." does not exclude the existence of another identical element in the process, method, article or terminal device comprising the said element.

[0089] The embodiments of the present invention have been described above in conjunction with the accompanying drawings, but the present invention is not limited to the above specific embodiments. The above specific embodiments are merely illustrative and not restrictive. Under the inspiration of the present invention, those of ordinary skill in the art can also make many forms without departing from the purpose of the present invention and the scope protected by the claims, and all of these are within the protection scope of the present invention.

Claims

1. A load demand interval probability prediction method considering abnormal meteorological conditions, characterized in that The load demand interval probability prediction method includes: Obtain multi-dimensional data and perform data fusion processing to obtain a multi-dimensional feature set. The multi-dimensional data includes: historical meteorological data; Use the principal component analysis method to reduce the dimension of the multi-dimensional feature set, and extract all meteorological factors in the historical meteorological data that affect the change of load demand under abnormal meteorological conditions; Analyze the combined features of all meteorological factors to obtain all extreme meteorological scenarios under abnormal meteorological conditions; Cluster all the extreme meteorological scenarios to obtain different types of extreme meteorological scenarios, and identify the key factors that affect the change of load demand under different types of extreme meteorological scenarios; Based on the key factors, construct time series models corresponding to different types of extreme meteorological scenarios. This time series model is used to predict the average value of load demand at any future time point; Adopt quantile regression combined with the time series model to construct an interval load prediction model. This interval load prediction model is used to predict the interval probability of load demand at any future time point.

2. The probability prediction method for load demand interval according to claim 1, wherein The multi-dimensional data further includes: historical load data, historical power grid operation status data; the historical meteorological data includes: abnormal meteorological data and normal meteorological data; Obtain multi-dimensional data and perform data fusion processing to obtain a multi-dimensional feature set, including: Adopt wavelet transform denoising and clustering anomaly detection methods to identify and remove outliers in the historical meteorological data, the historical load data, and the historical power grid operation status data to obtain correct data; Use time synchronization processing and data methods to fuse the correct data, match the historical load data, the historical power grid operation status data with the corresponding abnormal meteorological data or normal meteorological data, and obtain the correlation between the abnormal meteorological data or the normal meteorological data and the change of load demand at each time point; Label the correlation between the abnormal meteorological data and the change of load demand as data under abnormal meteorological conditions; Label the correlation between the normal meteorological data and the change of load demand as data under normal meteorological conditions.

3. The probability prediction method for load demand interval according to claim 1, wherein Use the principal component analysis method to reduce the dimension of the multi-dimensional feature set, and extract all meteorological factors in the historical meteorological data that affect the change of load demand under abnormal meteorological conditions, including: Through the principal component analysis method, transform the multi-dimensional feature set into a new low-dimensional feature space to obtain a low-dimensional feature set, so that the features corresponding to each data after dimension reduction are orthogonal to each other and do not contain duplicate information; Based on the low-dimensional feature set, calculate the non-linear correlation between the change of load demand and the historical meteorological data to obtain all the meteorological factors.

4. The load demand interval probability prediction method according to claim 3, wherein Based on the low-dimensional feature set, calculate the non-linear correlation between the change of load demand and the historical meteorological data to obtain all the meteorological factors, including: Adopt the correlation coefficient combined with non-linear metrics to calculate the non-linear correlation between the change of load demand and the historical meteorological data; Use Fourier transform to transform the non - linear correlation between the load demand change and the historical meteorological data from the time domain to the frequency domain, extract the periodic characteristics of each data, and identify all meteorological factors that have a non - linear impact on the load demand change under the abnormal meteorological conditions; Among them, the periodic characteristics are used as the step size when constructing the time series model.

5. The load demand interval probability prediction method according to claim 1, characterized in that, Analyze the combined characteristics of all meteorological factors to obtain all extreme meteorological scenarios under abnormal meteorological conditions, including: Construct combinations of all meteorological factors to obtain combined factors; Analyze and capture the interaction between the combined factors to obtain all the extreme meteorological scenarios.

6. The probability prediction method for load demand interval according to claim 1, wherein Cluster all the extreme meteorological scenarios to obtain different types of extreme meteorological scenarios, including: Cluster all the extreme meteorological scenarios using one or more of k - means, k - means++, k - medoids, k - median to obtain different types of extreme meteorological scenarios; For each type of extreme meteorological scenario, identify the key factors that affect the load demand change.

7. The probability prediction method for load demand interval according to claim 5, wherein The different types of extreme meteorological scenarios include: extremely hot meteorological scenarios and extremely cold meteorological scenarios; The combined factors corresponding to the extremely hot meteorological scenario are: The combined factors corresponding to the extremely cold meteorological scenario are: In the above two equations, represents the heat index, represents the cold index, represents the Fahrenheit temperature, represents the wind speed, represents the relative humidity, represents a constant; when the heat index exceeds the heat threshold, it is determined as the extremely hot meteorological scenario; when the cold index exceeds the cold threshold, it is determined as the extremely cold meteorological scenario.

8. The probability prediction method for load demand interval according to claim 4, characterized in that, Construct time series models corresponding to different types of extreme meteorological scenarios based on the key factors, including: For each type of extreme meteorological scenario, input its corresponding key factors, the historical power load data extracted from the historical load data, and the static characteristic data extracted from the historical power grid operation status data into the time series model according to the step size, and use a long - short - term memory network combined with an attention mechanism to construct a time series model corresponding to each type of extreme meteorological scenario; Among them, the attention mechanism is a multi - head parallel attention mechanism; the steps of constructing the time series model by combining the long - short - term memory network and the attention mechanism include: Select a vector sequence input to the attention mechanism from the time - series feature data at the step size, and each vector in the vector sequence is a d - dimensional vector containing the key factor, its corresponding historical power load data, and static characteristic data; Generate corresponding query vectors, key vectors, and value vectors for each vector; Calculate the respective attention scores according to the query vectors and key vectors of each vector in the vector sequence of each head in the multi - head parallel attention mechanism; Convert the respective attention scores and value vectors of each vector into a weight distribution through Softmax to obtain the output of each head; Concatenate the outputs of all heads to form a high - dimensional output vector; Use the high - dimensional output vector as the input of the long - short - term memory network to form the time series model.

9. The load demand interval probability prediction method according to claim 1, characterized in that Construct an interval load forecasting model by combining quantile regression with the time series model, including: At the output level of the time series model corresponding to each type of extreme meteorological scenario, the quantile regression method is introduced to expand the average value output at the output level into an interval probability prediction, and an interval load prediction model corresponding to each type of extreme meteorological scenario is constructed.

10. The probability prediction method for load demand interval according to claim 1, wherein After constructing the interval load prediction model, it further includes: Optimizing the interval load prediction model by using the historical meteorological data and historical load data.

Images