Quantile probability power load forecasting method and system

Abstract

The present invention discloses a quantile probability power load forecasting method and system, belonging to the technical field of power load forecasting. The method comprises screening the feature types of external feature data, removing the feature types whose influence on the power load is less than the relationship threshold, and the remaining feature types related to the power load; using the external feature data corresponding to the feature types related to the power load as a data set to train the QRGSNP model; inputting the external feature data corresponding to the feature types related to the power load at the time point to be predicted into the trained QRGSNP model to obtain the predicted values ​​of several different probabilities at the time point; using the predicted values ​​of several different probabilities at the time point as quantiles, and using different quantiles as upper and lower bounds to set confidence intervals, thereby performing interval prediction of the power load at the time point. The present invention uses the LASSO model to screen out the features with less relationship with the load in the data, and uses the QRGSNP model to perform quantile prediction, which significantly improves the prediction performance.

Description

Technical Field

[0001] The present invention belongs to the technical field of power load forecasting, and in particular relates to a quantile probability power load forecasting method and system. Background Art

[0002] In the operation and management of power systems, accurate load forecasting can help power companies optimize power generation and grid operation, and improve energy utilization. With the increasing popularity of renewable energy in smart grids and the continuous increase in building power load, its accurate prediction is crucial to system design, control and related optimization. So far, probabilistic forecasting methods have attracted more and more attention because they can evaluate the impact of various uncertainties. Probabilistic forecasting can be roughly divided into three categories: quantile forecasting, interval forecasting and probability density forecasting. Among them, the probabilistic forecasting method of quantile regression is more popular and develops rapidly. Therefore, how to improve the performance of quantile regression probabilistic load forecasting has become a research hotspot.

[0003] The existing probabilistic forecasting methods based on quantile regression are divided into two categories, namely, statistical-based methods and machine learning-based methods. Since machine learning can better handle nonlinear problems and multidimensional data, the forecasting performance is significantly improved compared with statistical methods. At present, the main machine learning algorithms used are: support vector machine, extreme gradient boosting, random forest and deep learning algorithms. Deep learning has automatic feature extraction and powerful nonlinear mapping capabilities, and can show excellent performance when processing complex data and tasks. Among them, quantile regression neural network (QRNN), long short-term memory network (LSTM) and gated recurrent unit network (GRU) models have been deeply studied in probabilistic load forecasting.

[0004] For example, the prior art Chinese patent application CN201811424684.2 discloses a method for power load forecasting. First, the DBSCAN algorithm is used to cluster the power load data to obtain the clustering result C and the class center; then the load characteristic data of the predicted day N days is input into the front-stage LSTM neural network as training data, and the predicted daily load characteristics are output; finally, the Euclidean distance between the predicted daily load characteristics and the class center after DBSACN clustering is calculated, and the load data corresponding to the class center with the smallest distance is merged with the original load data as training data and input into the back-stage LSTM neural network to output the predicted daily load curve.

[0005] The above-mentioned prior art uses LSTM neural network to predict power load. The above method is difficult to maintain prediction accuracy when processing long time series data. This is mainly because the traditional neural network model has information loss problems when capturing long-term dependencies, resulting in reduced prediction accuracy.

[0006] In addition, when selecting load data, if irrelevant features are input into the model, the model will be overfitted, which will reduce its prediction accuracy on new data. Summary of the invention

[0007] In view of this, the present invention provides a quantile probability power load forecasting method and system, which can screen out features with less correlation with power load in data and reduce the impact of information loss on forecasting accuracy.

[0008] In order to solve the above technical problems, the technical solution of the present invention is to adopt a quantile probability power load forecasting method, comprising:

[0009] Collect samples, each sample includes power load data and corresponding external characteristic data;

[0010] Screen the feature types of the external feature data, remove the feature types whose impact on the power load is less than the relationship threshold, and the remaining feature types are related to the power load;

[0011] The external feature data corresponding to the feature types related to power load are used as the data set to train the QRGSNP model;

[0012] The external feature data corresponding to the feature type related to the power load at the time point to be predicted is input into the trained QRGSNP model to obtain the predicted values ​​with different probabilities at the time point;

[0013] The predicted values ​​of several different probabilities at the time point are taken as quantiles, and different quantiles are used as upper and lower bounds to set the confidence interval, so as to make an interval prediction of the power load at the time point.

[0014] As an improvement, the LASSO model is used to screen the external feature data;

[0015] The LASSO model is:

[0016] ;

[0017] Where n is the number of samples; p is the number of feature types; y i is the actual value of the ith power load; is the predicted value of the ith power load; β j is the j-th feature model parameter; π is the weight of the L1 regularization term.

[0018] As a further improvement, the value of π is 0.001.

[0019] As another further improvement, the QRGSNP model is formed by rewriting the loss function of the GSNP model;

[0020] The rewritten loss function is:

[0021] ;

[0022] Where F is the loss value; Q is the number of quantiles; is in quantile The ith power load forecast value under i is the actual value of the ith power load.

[0023] As an improvement, methods for setting confidence intervals using different quantiles as upper and lower bounds include:

[0024] The QRGSNP model outputs 99 predicted values ​​with different probabilities as quantiles, and the probabilities of the 99 predicted values ​​are an arithmetic progression from 1% to 99%;

[0025] The 5% probability quantile and the 95% probability quantile were selected as the upper and lower limits of the 90% confidence interval; the 7% probability quantile and the 92% probability quantile were selected as the upper and lower limits of the 85% confidence interval; and the 10% probability quantile and the 90% probability quantile were selected as the upper and lower limits of the 80% confidence interval.

[0026] As an improvement, the samples are preprocessed after collection, including screening outliers, filling missing values, and normalization;

[0027] After filtering out the outliers, they are marked as missing values, and the median is used to fill in the missing values.

[0028] As an improvement, the method for screening outliers includes:

[0029] After sorting the data, three quantile points are divided; the quantile points are 25% probability quantile, median and 75% quantile;

[0030] If the data is not in the interval [Q1-1.5×IQR, Q3-1.5×IQR], the data is an outlier; where IQR=Q3-Q1, Q3 is the 75th quantile, and Q1 is the 25th quantile.

[0031] The present invention also provides a quantile probability power load forecasting system, comprising:

[0032] A sample collection module, used to collect samples, each sample includes power load data and corresponding external characteristic data;

[0033] A feature type screening module is used to screen the feature types of external feature data, remove feature types whose impact on power load is less than a relationship threshold, and the remaining feature types are related to power load;

[0034] A model training module, used to train the QRGSNP model using external feature data corresponding to feature types related to power load as a data set;

[0035] The prediction value output module is used to input the external feature data corresponding to the feature type related to the power load at the time point to be predicted into the trained QRGSNP model to obtain the prediction values ​​of several different probabilities at the time point;

[0036] The interval prediction module is used to take the predicted values ​​of several different probabilities at the time point as quantiles, and use different quantiles as upper and lower bounds to set the confidence interval, so as to perform interval prediction of the power load at the time point.

[0037] As an improvement, the feature type screening module uses the LASSO model to screen the external feature data;

[0038] The LASSO model is:

[0039] ;

[0040] Where n is the number of samples; p is the number of feature types; y i is the actual value of the ith power load; is the predicted value of the ith power load; β j is the j-th feature model parameter; π is the weight of the L1 regularization term.

[0041] As an improvement, the QRGSNP model is formed by rewriting the loss function of the GSNP model;

[0042] The rewritten loss function is:

[0043] ;

[0044] Where F is the loss value; Q is the number of quantiles; is in quantile The ith power load forecast value under i is the actual value of the ith power load.

[0045] The present invention is beneficial in that:

[0046] The present invention adopts LASSO model to screen out the features with less correlation with the prediction results in the data, thereby streamlining the model and ensuring the accuracy of the prediction results. LASSO algorithm is a powerful linear regression technique, which realizes feature selection and parameter estimation by introducing L1 regularization term. Its main advantage is that it can automatically select important features effectively, reduce the coefficients of features with weak relationship with the prediction target to zero, thereby simplifying the model structure and improving its interpretability. This makes LASSO particularly suitable for processing high-dimensional data, especially when the number of features is greater than the number of samples, it can effectively avoid overfitting and improve the generalization ability of the model. In addition, LASSO shows certain robustness to outliers because it reduces the sensitivity of the model to extreme values ​​by regularization. The purpose of LASSO is to improve prediction accuracy by reducing model complexity, help users identify the features that have the greatest impact on the response variable, and still provide reliable estimates when there is a problem of multicollinearity in the data.

[0047] In addition, the present invention modifies the GSNP system so that it has the function of quantile prediction. Due to its gating mechanism, the traditional GSNP system is particularly suitable for processing time series data prediction, such as power load forecasting and financial data analysis, which require processing time dependencies. However, the traditional GSNP can only generate a single prediction value and cannot quantify uncertainty. By modifying the loss function, the GSNP system is transformed into a model with probabilistic prediction capabilities, and combined with the LASSO algorithm, the prediction performance is significantly improved. In response to the challenges of the volatility of renewable energy and the continuous increase in building power load, the new GSNP model provides a probabilistic distribution of load demand, which provides strong support for power companies to better manage and integrate these unstable resources, thereby improving the overall efficiency and stability of the power grid. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, each element or part is not necessarily drawn according to the actual scale. Obviously, the drawings described below are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can also be obtained based on these drawings without paying creative labor.

[0049] Figure 1 It is a flow chart of the present invention.

[0050] Figure 2 This is the structural schematic diagram of the QRGSNP model.

[0051] Figure 3(a) is a schematic diagram of the 80% confidence interval prediction results.

[0052] Figure 3(b) is a schematic diagram of the 85% confidence interval prediction results.

[0053] Figure 3(c) is a schematic diagram of the 90% confidence interval prediction results.

[0054] Figure 4(a) is a schematic diagram of the comparison of AIS indicator results.

[0055] Figure 4(b) is a schematic diagram for comparing PL indicator results.

[0056] Figure 5 Schematic diagram of the deterministic point prediction results of different models. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0058] Herein, suffixes such as "module", "component" or "unit" used to represent elements are only used to facilitate the description of the present invention, and have no specific meanings by themselves. Therefore, "module", "component" or "unit" can be used mixedly.

[0059] In this document, the terms "upper", "lower", "inner", "outer", "front", "back", "one end", "the other end" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a specific orientation, be constructed and operate in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first" and "second" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance.

[0060] In this document, unless otherwise clearly specified and limited, the terms "installed", "provided with", "connected", etc. should be understood in a broad sense. For example, "connected" can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection, a direct connection, or an indirect connection through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0061] Herein "and / or" includes any and all combinations of one or more of the associated listed items.

[0062] Herein, "plurality" means two or more than two, ie, it includes two, three, four, five, etc.

[0063] Example 1

[0064] like Figure 1 As shown, the present invention provides a quantile probability power load forecasting method, comprising:

[0065] S1 collects samples, each sample includes power load data and corresponding external characteristic data.

[0066] In this embodiment, the external characteristics of the sample include maximum temperature, minimum temperature, average temperature, average humidity, rainfall, and of course any other external characteristic data that is considered to affect the regional power load.

[0067] In fact, external characteristics can be divided into socio-economic factors, environmental factors, meteorological factors, policy factors and other technical factors. The following are the detailed external characteristics that affect power load:

[0068] 1. Socioeconomic factors

[0069] Population factors: An increase in population will lead to an increase in electricity demand, especially for residential electricity consumption.

[0070] Economic development level: With the improvement of economic development level, the demand for industrial and commercial electricity will increase, and the increase in people's income will also lead to an increase in the demand for domestic electricity.

[0071] Industrial structure: Different industries have different electricity demand intensities. The increase in high-energy-consuming industries will significantly increase the electricity load.

[0072] 2. Environmental factors

[0073] Urbanization: The accelerated urbanization process will lead to an increase in urban population, thereby increasing residential and commercial electricity load.

[0074] Changes in living environment: The addition of new buildings and the upgrading of electrical equipment in existing buildings will increase the power load.

[0075] 3. Meteorological factors

[0076] Seasonal variation: Electricity demand is higher in summer and winter because it is used for air conditioning in summer and for heating in winter.

[0077] Weather conditions: Extreme weather conditions (such as heat waves and cold snaps) can significantly increase electricity loads.

[0078] Sunshine hours: Increase or decrease in sunshine hours will affect the efficiency of the solar power generation system, thereby affecting the power load.

[0079] 4. Policy factors

[0080] Energy policy: The government’s energy policy, energy conservation and emission reduction policy, and renewable energy policy will affect the structure and changes of power load.

[0081] Price policy: Adjustments to electricity prices will directly affect user behavior, which in turn affects electricity load.

[0082] Grid management policies: Grid dispatch and load management policies also affect the distribution and changes of power load.

[0083] 5. Other technical factors

[0084] Popularization of electrical equipment: The popularization and replacement of household electrical appliances will increase the demand for electricity.

[0085] The application of new technologies: such as the popularization of electric vehicles and the application of smart home technologies will increase the power load.

[0086] Grid upgrade: The intelligent upgrade of the grid and the application of distributed energy systems will change the distribution and characteristics of power loads.

[0087] When collecting samples, the above features can all be collected as external feature data.

[0088] S2 preprocesses the samples, including screening outliers, filling missing values, and normalization.

[0089] If there are missing values ​​or outliers in the collected data, it will affect the subsequent prediction. Therefore, it is necessary to preprocess the data, remove outliers and fill in missing values.

[0090] In this embodiment, the quartile method is used to screen outliers, and the specific steps include:

[0091] S21 sorts the data and divides them into three quantile points; the quantile points are 25% probability quantile, median and 75% quantile;

[0092] S22 If the data is not in the interval [Q1-1.5×IQR,Q3-1.5×IQR], the data is an outlier; where IQR=Q3-Q1, Q3 is the 75th quantile, and Q1 is the 25th quantile.

[0093] After the outliers are screened out by the above method, they are marked as missing values, and the median is used to fill the missing values. Median is a concept in statistics, which represents the value in the middle of a set of data. Specifically, the median is the number in the middle after a set of data is arranged in order of numerical size. If the number of data is an odd number, the median is the number in the middle; if the number of data is an even number, the median is the average of the two middle numbers.

[0094] Filling missing values ​​with the median can avoid affecting the prediction results.

[0095] In addition, the features in the data have different magnitudes, which leads to a decrease in the accuracy of the model when predicting. For this reason, normalization is performed to scale the features to the range of 0-1. The formula is as follows:

[0096] ;

[0097] Where X is the feature to be scaled; X min is the minimum value of this feature in the data; X max is the maximum value of this feature in the data; X s is the scaled feature.

[0098] S3 screens the feature types of the external feature data, removes the feature types whose impact on the power load is less than the relationship threshold, and the remaining feature types are related to the power load.

[0099] In the present invention, the LASSO model is used to screen the external feature data;

[0100] The LASSO model is:

[0101] ;

[0102] Where n is the number of samples; p is the number of feature types, which is 5 in this implementation; y i is the actual value of the ith power load; is the predicted value of the ith power load; β j is the jth feature model parameter; π is the weight of the L1 regularization term. L1 regularization (L1 regularization or lasso) is an important method in machine learning. In the process of support vector machine learning, it is actually a process of finding the optimal solution to the cost function. Therefore, L1 regularization adds the L1 norm to the cost function to make the learning result meet the sparseness requirements, thus facilitating people to extract features.

[0103] In this implementation, LASSO feature selection is performed on the pre-processed external feature data such as minimum temperature, maximum temperature, average temperature, average humidity and rainfall. The L1 regularization term changes some parameters in the model to 0 during the optimization process, thereby achieving feature selection. Larger values ​​will compress more coefficients to zero, thus selecting fewer features. The value will retain more features. This article will determine the weight of L1 regularization Set to 0.001.

[0104] The LASSO model will output a correlation coefficient graph for each feature. In this implementation, features with a correlation coefficient of 0 are removed to ensure the simplicity and accuracy of the subsequent prediction model.

[0105] S4 uses the external feature data corresponding to the feature types related to power load as a data set to train the QRGSNP model.

[0106] The GSNP model is implemented under the framework of the Recurrent Neural Network (RNN), so GSNP can also be regarded as a recurrent neural network that introduces a neuron pulse mechanism. It has two control gates, which control the update of the gated neuron state and can selectively retain or forget information in long-term load forecasting. In addition, compared with the previous information transmission method of recurrent neural networks, GSNP uses discrete pulse signals to transmit between neurons, and does not need to continuously activate all neurons, so that important information can be better retained in long time series, thereby further improving the performance of the model.

[0107] The GSNP model consists of a gated neuron The GSNP system is composed of. At each moment, the neuron has an external input, and each neuron calculates the current output based on the current external input and the state of the previous neuron. GSNP can be regarded as a directed graph, where each node is a neuron and is gated, and each line represents a synaptic connection between two neurons.

[0108] The Pulse Rule The pulse condition can be expressed as , where T is the pulse threshold; is the gating neuron where the rule is located In time state; a is a pulse signal. If exist If the following is established, then will be triggered. The pulse is consumed, The impulse is generated and passed to the next neuron. Therefore, these two impulse functions are called consumption function and generation function. Its state equation is expressed as:

[0109]

[0110] in, ; For time External input; and The reset gate and the consumption gate control the update of the neuron state respectively, and both gating functions are nonlinear. In addition, the output equation is expressed as:

[0111]

[0112] In long-term load forecasting, the GSNP system only considers one neuron, denoted as , and two special assumptions are considered in the prediction.

[0113] 1) : Neurons can be triggered at any time.

[0114] 2) : Let the consumption function be equal to the generation function.

[0115] Only one neuron is considered in the system, so the state equation and output equation can be rewritten as:

[0116]

[0117] The state of the neuron and The GSNP model is updated under the control of the two gates, so it is suitable for dealing with long time series load forecasting problems.

[0118] In order to be more general, the model is extended to multiple dimensions and recorded as , and . Its output equation and state equation are rewritten as:

[0119]

[0120] in, , , , and is the weight matrix; and is the bias vector.

[0121] Furthermore, the gating equation can be written as:

[0122] 1) Reset gate

[0123]

[0124] 2) Consumption gate

[0125]

[0126] in, , , and is the weight matrix; and is the bias vector; is a non-linear activation function.

[0127] Due to its gating mechanism, the GSNP system is particularly suitable for processing time series data forecasts, such as power load forecasting and financial data analysis, which require processing time dependencies. However, GSNP can only generate a single forecast value and cannot quantify uncertainty.

[0128] In the present invention, by modifying the loss function, the GSNP system is transformed into a model with probability prediction capability. Specifically, the loss function of the existing GSNP model in the present invention is changed to a quantile regression loss function, thereby constructing a GSNP quantile regression model (Quantile-Regression-GSNP, QRGSNP), such as Figure 2 shown.

[0129] More specifically, the quantile regression loss function is:

[0130] ;

[0131] Where F is the loss value; Q is the number of quantiles; is in quantile The ith power load forecast value under i is the actual value of the ith power load.

[0132] The feature data selected in the previous step is input into the QRGSNP model to train the model, so that the model can output several power load forecast values ​​with different probabilities, unlike the traditional GSNP model which can only output a single power load forecast value.

[0133] S5 inputs the external feature data corresponding to the feature type related to the power load at the time point to be predicted into the trained QRGSNP model to obtain the predicted values ​​of several different probabilities at the time point.

[0134] To apply the QRGSNP model to the prediction of power load, we only need to input the external characteristic data at the time point to be predicted. The external characteristic data is screened by the LASSO model, and its characteristic type is highly correlated with the final prediction result.

[0135] In this embodiment, by setting (setting the number of quantiles Q in the loss function to 99), the model can output 99 power load forecast values ​​with different probabilities. For the convenience of calculation, the probability of the 99 power load forecast values ​​with different probabilities is an arithmetic progression from 1% to 99%, that is, from 1%, 2%, 3%...99%.

[0136] S6 uses the predicted values ​​of several different probabilities at the time point as quantiles, and uses different quantiles as upper and lower bounds to set confidence intervals, thereby performing interval prediction of the power load at the time point.

[0137] In this embodiment, the QRGSNP model outputs 99 prediction values ​​with different probabilities as quantiles, and the probabilities of the 99 prediction values ​​are an arithmetic progression from 1% to 99%.

[0138] The so-called quantile is a method used in statistics to describe the distribution of data. It divides a data set into several parts to better understand the distribution of the data. Common quantiles are:

[0139] Quartiles: Divide the data into four parts, each containing 25% of the data.

[0140] Deciles: Divide the data into ten parts, each containing 10% of the data.

[0141] Percentiles: Divide the data into 100 parts, each containing 1% of the data. In this embodiment, 99 prediction values ​​with different probabilities are output, the purpose of which is to form percentiles.

[0142] More specifically in this implementation, the 5% probability quantile and the 95% probability quantile are selected as the upper and lower limits of the 90% confidence interval; the 7% probability quantile and the 92% probability quantile are selected as the upper and lower limits of the 85% confidence interval; and the 10% probability quantile and the 90% probability quantile are selected as the upper and lower limits of the 80% confidence interval.

[0143] The so-called 90% confidence interval can be understood as the probability that the power load will fall within this interval. The same goes for the 85% and 80% confidence intervals.

[0144] Of course, in other embodiments, other numbers of predicted values ​​may be output, for example, three predicted values ​​may be output, with probabilities of 25%, 50%, and 75%, respectively. The upper and lower limits of the confidence interval may also be selected from 75% and 25% quantiles, thereby forming a 50% confidence interval.

[0145] S7 model performance evaluation.

[0146] The present invention adopts four indicators, namely, Pinball Loss (PL), interval coverage probability (PICP), interval average bandwidth (PINAW) and average interval score (Average Interval Score, AIS), to evaluate the load forecasting interval performance.

[0147] The specific case uses the 2016 Electrician Competition Cup data set, with a time interval of 60 minutes. The data set contains four years of load data, and its feature types are minimum temperature, maximum temperature, average temperature, average humidity and rainfall. A total of 26,545 samples.

[0148] First, the correlation between each feature and load was generated based on LASSO. The minimum temperature had the largest correlation, followed by the average temperature and the maximum temperature. The average humidity and rainfall had a correlation coefficient of 0 and were screened out. Therefore, the three feature data of minimum temperature, average temperature and maximum temperature were used as input.

[0149] 1) Interval prediction results

[0150] The prediction results of the proposed method at 80%, 85% and 90% confidence intervals are shown in Figure 3(a), Figure 3(b) and Figure 3(c) respectively. It can be seen that even in the stage of large load fluctuations, the prediction interval can well cover the true value, and the upper and lower limits of the interval prediction have the same trend as the curve of the true value.

[0151] The comparison diagrams of PL and AIS with other models are shown in Figure 4(a) and Figure 4(b), and the specific evaluation data are shown in Tables 1 and 2.

[0152] Table 1 Prediction indicators of each model interval

[0153]

[0154] Table 2 PL indicators of each model

[0155]

[0156] It can be seen that in the 90% confidence interval and the 80% confidence interval, except for LSTM, the prediction interval coverage probability of other models exceeds the corresponding PINC, so the prediction results of LSTM in these two confidence intervals are unreliable. Among them, the interval width of the method proposed in this paper is the narrowest among all models while ensuring a certain interval coverage rate. Under the three confidence intervals, the average value of AIS of the method in this paper is improved by 25%, 26.56%, 56.89%, 78.25% and 82.95% compared with GSNP, LSTM, GRU, CNN and RNN, respectively, and the average PL error is reduced by 10.48%, 18.1%, 36.65%, 71.47% and 76.34%, respectively. These results can prove the advantages of this method in interval prediction and quantile prediction.

[0157] 2) Median prediction results

[0158] The median prediction results of this method and the comparison with other models are as follows: Figure 5 ,From the figure, we can see that the proposed model has higher ,accuracy than other models in deterministic point prediction.,Table 3 shows the evaluation indicators of various deterministic ,point prediction.

[0159] Table 3 Comparison of indicators of deterministic point prediction models

[0160]

[0161] The proposed method has the smallest error in MAPE and MAE, and the highest R2. These results prove that the proposed model has superior performance in deterministic point prediction.

[0162] Example 2

[0163] The present invention also provides a quantile probability power load forecasting system, comprising:

[0164] A sample collection module, used to collect samples, each sample includes power load data and corresponding external characteristic data;

[0165] A feature type screening module is used to screen the feature types of external feature data, remove feature types whose impact on power load is less than the relationship threshold, and leave the remaining feature types related to power load; the feature type screening module uses the LASSO model to screen the external feature data;

[0166] The LASSO model is:

[0167] ;

[0168] Where n is the number of samples; p is the number of feature types; y i is the actual value of the ith power load; is the predicted value of the ith power load; β j is the j-th feature model parameter; π is the weight of the L1 regularization term.

[0169] A model training module, used for training a QRGSNP model using external feature data corresponding to feature types related to power load as a data set; the QRGSNP model is formed by rewriting the loss function of the GSNP model;

[0170] The rewritten loss function is:

[0171] ;

[0172] Where F is the loss value; Q is the number of quantiles; is in quantile The ith power load forecast value under i is the actual value of the ith power load.

[0173] The prediction value output module is used to input the external feature data corresponding to the feature type related to the power load at the time point to be predicted into the trained QRGSNP model to obtain the prediction values ​​of several different probabilities at the time point;

[0174] The interval prediction module is used to take the predicted values ​​of several different probabilities at the time point as quantiles, and use different quantiles as upper and lower bounds to set the confidence interval, so as to perform interval prediction of the power load at the time point.

[0175] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or device including the element.

[0176] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, a magnetic disk, or an optical disk), and includes a number of instructions for a computer terminal (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in each embodiment of the present invention.

[0177] The above are only preferred embodiments of the present invention. It should be noted that the above preferred embodiments should not be regarded as limiting the present invention, and the protection scope of the present invention should be based on the scope defined by the claims. For ordinary technicians in this technical field, several improvements and modifications can be made without departing from the spirit and scope of the present invention, and these improvements and modifications should also be regarded as the protection scope of the present invention.

Claims

1. A quantile probability power load forecasting method, characterized in that include: Collect samples, each sample including power load data and corresponding external characteristic data; Screen the feature types of the external feature data, remove the feature types whose impact on the power load is less than the relationship threshold, and the remaining feature types are related to the power load; Using external feature data corresponding to feature types related to power load as a data set to train a QRGSNP model; the QRGSNP model is formed by rewriting the loss function of the GSNP model; The rewritten loss function is: ； Where F is the loss value; Q is the number of quantiles; is in quantile The ith power load forecast value under i is the actual value of the ith power load; n is the sample size; The external feature data corresponding to the feature type related to the power load at the time point to be predicted is input into the trained QRGSNP model to obtain the predicted values ​​with different probabilities at the time point; The predicted values ​​of several different probabilities at the time point are taken as quantiles, and different quantiles are used as upper and lower bounds to set the confidence interval, so as to make an interval prediction of the power load at the time point.

2. A quantile probability power load forecasting method according to claim 1, characterized in that: Use LASSO model to screen external feature data; The LASSO model is: ； Where n is the number of samples; p is the number of feature types; y i is the actual value of the ith power load; is the predicted value of the ith power load; β j is the j-th feature model parameter; π is the weight of the L1 regularization term.

3. A quantile probability power load forecasting method according to claim 2, characterized in that: The value of π is 0.

001.

4. A quantile probability power load forecasting method according to claim 1, characterized in that Methods for setting confidence intervals using different quantiles as upper and lower bounds include: The QRGSNP model outputs 99 predicted values ​​with different probabilities as quantiles, and the probabilities of the 99 predicted values ​​are an arithmetic progression from 1% to 99%; The 5% probability quantile and the 95% probability quantile were selected as the upper and lower limits of the 90% confidence interval; the 7% probability quantile and the 92% probability quantile were selected as the upper and lower limits of the 85% confidence interval; and the 10% probability quantile and the 90% probability quantile were selected as the upper and lower limits of the 80% confidence interval.

5. A quantile probability power load forecasting method according to claim 1, characterized in that: After collecting samples, the samples are preprocessed, including screening outliers, filling missing values, and normalization; After filtering out the outliers, they are marked as missing values, and the median is used to fill in the missing values.

6. A quantile probability power load forecasting method according to claim 5, characterized in that The method for screening outliers comprises: After sorting the data, three quantile points are divided; the quantile points are 25% probability quantile, median and 75% quantile; If the data is not in the interval [Q1-1.5×IQR, Q3-1.5×IQR], the data is an outlier; where IQR=Q3-Q1, Q3 is the 75th quantile, and Q1 is the 25th quantile.

7. A quantile probability power load forecasting system, characterized in that include: A sample collection module, used to collect samples, each sample includes power load data and corresponding external characteristic data; A feature type screening module is used to screen the feature types of external feature data, remove feature types whose impact on power load is less than a relationship threshold, and the remaining feature types are related to power load; A model training module, used for training a QRGSNP model using external feature data corresponding to feature types related to power load as a data set; the QRGSNP model is formed by rewriting the loss function of the GSNP model; The rewritten loss function is: ； Where F is the loss value; Q is the number of quantiles; is in quantile The ith power load forecast value under i is the actual value of the ith power load; n is the number of samples; The prediction value output module is used to input the external feature data corresponding to the feature type related to the power load at the time point to be predicted into the trained QRGSNP model to obtain the prediction values ​​of several different probabilities at the time point; The interval prediction module is used to take the predicted values ​​of several different probabilities at the time point as quantiles, and use different quantiles as upper and lower bounds to set the confidence interval, so as to perform interval prediction of the power load at the time point.

8. A quantile probability power load forecasting system according to claim 7, characterized in that: The feature type screening module uses the LASSO model to screen the external feature data; The LASSO model is: ； Where n is the number of samples; p is the number of feature types; yi is the actual value of power load; is the power load forecast value; β j is the j-th feature model parameter; π is the weight of the L1 regularization term.

Images