Data-driven power similar day screening method and system

Abstract

The invention relates to the technical field of time series and power load prediction and data analysis, in particular to a data-driven power similar day screening method and system, and the method comprises the steps: carrying out the estimation through employing a first principle, and building a preliminary basis for similar day screening; constructing a model and training to learn a first weight of load similarity; thirdly, the obtained first weight is further processed, and weight distribution is corrected through methods such as moving average and linear regression; power load prediction is carried out on the basis of the processed first weight, the data-driven similar day screening method effectively reduces dependence on artificial experience, the universality and adaptability of the model are improved, and micro modes and complex relations which are difficult to perceive manually are recognized in a data-driven mode; therefore, the screened similar day can reflect the actual power utilization mode change more accurately, the accuracy and reliability of load prediction are greatly improved, and more scientific support is provided for stable operation of a power system.

Description

Technical Field

[0001] The present invention relates to the technical field of time series, power load forecasting and data analysis, and in particular to a data-driven power similar day screening method and system. Background Art

[0002] Similar days play an important role in the day-ahead power load forecasting. They can provide strong data support for power workers to formulate power generation plans, thereby achieving more accurate load forecasting and enhancing the economy and stability of power system operation. At present, the screening of similar days still relies heavily on the experience and judgment of staff. This is because the power consumption pattern of a region is closely related to the laws of local social activities. For example, holidays, special adjustment days or large-scale public events will significantly affect the power consumption pattern. These social activities often contain rich cultural and customary elements, which are difficult to summarize with unified rules. The data-driven method provides an innovative solution for the screening of similar days. By applying machine learning technology, the intrinsic connection between regional power consumption patterns can be automatically captured, and a similarity discrimination model can be constructed to achieve automatic screening of similar days, thereby reducing dependence on manual experience. Summary of the invention

[0003] In view of the above problems existing in the prior art, the present invention is proposed.

[0004] Therefore, the technical problem to be solved by the present invention is: at the current stage, the screening of similar electricity days requires staff to rely on their own experience to make decisions based on calendar, weather, social activities and other conditions, mainly to improve the automation and accuracy of the screening of similar electricity days; by introducing a data-driven approach, an intelligent model is established to objectively evaluate and predict the similarity of electricity demand, reduce the impact of human factors, and improve decision-making efficiency and accuracy.

[0005] To solve the above technical problems, the present invention provides the following technical solutions, a data-driven method for screening similar electricity days, comprising: using the first principle to make an estimate and establish a preliminary screening basis; constructing a first model and training it to learn the first weight; processing the first weight; and making a prediction based on the processed first weight.

[0006] As a preferred solution of the data-driven power similar day screening method described in the present invention, the establishment of a preliminary screening basis includes making a preliminary estimate of historical data based on first principles to provide a basic basis for subsequent screening.

[0007] As a preferred solution of the data-driven electricity similar day screening method described in the present invention, the construction of the first model and training includes using a neural network to construct the first model, learning the first weight distribution through a training process, and improving the accuracy of screening.

[0008] As a preferred solution of the data-driven power similar day screening method described in the present invention, the processing of the first weight includes using a class of methods to preliminarily calculate the first weight and make corrections to ensure the accuracy of the first weight.

[0009] As a preferred solution of the data-driven electricity similar day screening method described in the present invention, the prediction includes sorting the historical data using the modified first weight and selecting the data, thereby laying a foundation for accurate prediction.

[0010] As a preferred scheme of the data-driven electricity similar day screening method described in the present invention, wherein: the first principle includes the Kalman filtering principle, which estimates the similarity of electricity consumption patterns on different dates through basic processing of historical load data, providing a reliable basic basis for screening similar days; Kalman filtering can perform minimum variance estimation on historical data to obtain the first weight of the actual electricity consumption pattern.

[0011] As a preferred solution of the data-driven power similar day screening method described in the present invention, the first weight includes a similarity weight, which is used to measure the similarity of power consumption patterns between different historical dates and the day to be predicted; the similarity weight is obtained by comprehensive calculation of the Kalman filter principle and the neural network model, and is corrected in the subsequent processing stage to ensure the accuracy of the screened similar days;

[0012] The processing of the first weight also includes performing sliding average and linear regression correction to eliminate irrelevant components.

[0013] Another object of the present invention is to provide a data-driven power similar day screening system, which accurately estimates load similarity by applying Kalman filtering principles and neural network technology; the system consists of four parts: a preliminary screening module, a model training module, a weight processing module and a prediction execution module: the preliminary screening module uses Kalman filtering to estimate the load pattern of historical data; the model training module learns the weights of load similarity based on a neural network; the weight processing module corrects the weight distribution through sliding average and linear regression to eliminate irrelevant factors; finally, the prediction execution module screens similar days and predicts power load based on the processed weights.

[0014] In order to solve the above technical problems, the present invention provides the following technical solutions: a data-driven power similar day screening system, comprising: a preliminary screening module, a model training module, a weight processing module and a prediction execution module;

[0015] The preliminary screening module uses first principles to make estimates and establish a preliminary screening basis;

[0016] The model training module constructs and trains a first model to learn a first weight;

[0017] The weight processing module processes the first weight;

[0018] The prediction execution module performs prediction based on the processed first weight.

[0019] A computer device includes a memory and a processor, wherein the memory stores a computer program, and is characterized in that when the processor executes the computer program, the steps of the data-driven power similarity day screening method as described above are implemented.

[0020] A computer-readable storage medium stores a computer program thereon, wherein when the computer program is executed by a processor, the steps of the data-driven power similarity day screening method as described above are implemented.

[0021] The beneficial effects of the present invention are as follows: on the one hand, the data-driven similar day screening method can effectively reduce the dependence on manual experience and show higher universality and adaptability; on the other hand, the data-driven method can capture those subtle patterns and complex relationships that are difficult for humans to identify, thereby screening out more accurate similar days and improving the accuracy and reliability of power load forecasting. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0023] Figure 1 A flow chart of a data-driven power similar day screening method provided by an embodiment of the present invention.

[0024] Figure 2 An inventive scheme diagram of a data-driven power similar day screening method provided for one embodiment of the present invention.

[0025] Figure 3 A similarity weight estimation diagram based on a dual convolutional neural network for a data-driven electricity similar day screening method provided by an embodiment of the present invention.

[0026] Figure 4 A weight correction diagram based on sliding average and linear regression for a data-driven power similar day screening method provided by an embodiment of the present invention.

[0027] Figure 5 An experimental data diagram of a data-driven power similar day screening method provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0028] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0029] Example 1, reference Figure 1-Figure 2 , is an embodiment of the present invention, which provides a data-driven power similar day screening method, including:

[0030] S1: Use first principles to make estimates and establish a preliminary screening basis.

[0031] It should be noted that if Figure 1 S1 and Figure 2 As shown, establishing the preliminary screening basis includes making a preliminary estimate of the historical data based on the first principle, which provides a basic basis for subsequent screening; the first principle includes the Kalman filter principle, which estimates the similarity of power consumption patterns on different dates through basic processing of historical load data, and provides a reliable basis for screening similar days; the Kalman filter can perform minimum variance estimation on historical data and obtain the first weight of the actual power consumption pattern.

[0032] Furthermore, similar days have similar electricity consumption patterns. First, we assume that the daily electricity consumption in the region conforms to the following model:

[0033] L(t0=(L M +ε t 0L T (t) (1)

[0034] Among them, L(t) represents the historical load data of day t, L M Represents the power consumption mode, ε t represents the change in electricity consumption pattern caused by random factors on that day, L T (t) represents the daily electricity consumption of the region.

[0035] In reality, we can easily collect a large amount of historical electricity consumption data, which allows us to analyze the electricity consumption pattern L M Make an estimate:

[0036]

[0037] The existence of the error term ε means that the variance of the electricity consumption pattern estimated using different dates is different. If the error term follows a normal distribution, the electricity consumption pattern estimated using the historical data on the i-th day also follows a normal distribution:

[0038]

[0039] In the embodiment of the present application, the first principle is the Kalman filter principle, which estimates the similarity of electricity consumption patterns on different dates through basic processing of historical load data, providing a reliable basis for screening similar days; Kalman filtering can perform minimum variance estimation on historical load data, so that the screening of similar days is based on higher accuracy; specifically, Kalman filtering generates an estimated value of the load pattern by weighted fusion of estimates of multiple historical load data, and in the process performs weighted processing on the similarities of different dates; the weight is calculated based on the size of the estimated variance, and estimates with smaller variances are given higher weights, which means that the load pattern on that date is closer to the actual electricity consumption.

[0040] In an optional embodiment, the first principle can also be implemented in other ways, such as using a simple historical average method to estimate the similarity of load patterns on similar days. This method directly averages the historical load data and calculates the average power consumption on different dates as the similarity evaluation index, thereby screening out similar days with the closest load patterns. This method does not rely on Kalman filtering or neural network models, and only uses statistical averages to judge similarity. The calculation is simple and the implementation cost is low. However, this method has poor accuracy and robustness when dealing with complex load pattern changes, cannot effectively capture subtle differences between different dates, and is easily disturbed by short-term abnormal data, so the accuracy of the screening results is relatively low.

[0041] In an optional embodiment, the first principle can also be implemented in other ways, such as a similarity measurement method based on fixed rules, such as using linear interpolation to estimate historical load data and select similar days; in this method, the historical data is first interpolated according to specific rules, grouped by preset date features (such as weekdays, weekends, holidays), and then the similarity measurement in each group is calculated separately; in this way, load data of different dates will be classified under specific rules for similarity comparison; although the linear interpolation method is simpler than the Kalman filter and can achieve a certain similarity screening effect, it has poor adaptability to dynamically changing load patterns, and it is difficult to handle complex fluctuations in electricity demand. In addition, the accuracy is not high when facing multi-dimensional data, and the reliability of the screening results is low.

[0042] In the embodiment of the present application, the basic processing is weighted fusion, and the Kalman filter can perform minimum variance estimation on historical data to obtain the first weight of the actual power consumption pattern. Specifically, during the calculation process, the Kalman filter estimates the current load pattern through weighted fusion of historical load data, and uses weights to represent the similarity of load data on different dates; the allocation of these weights depends on the variance of each estimate. When the variance of the load estimate on a certain date is small, it means that the load pattern on that date is closer to the actual power consumption, so it is given a greater weight; this variance weighted method ensures that the date that best matches the current load pattern is selected in the historical data, thereby improving the accuracy of similar day screening.

[0043] In an optional embodiment, basic processing can also be implemented in other ways, such as directly using the moving average method to estimate the similarity of electricity consumption patterns on similar days; specifically, the moving average of a fixed window period of historical load data is used as an evaluation indicator of similarity; this method does not require complex weighted fusion, but only requires smoothing to obtain the average load value of each date, and then, based on the electricity consumption on the day to be predicted, the date with the closest load pattern is selected as a similar day; although the moving average method is simple to implement and has a low computational cost, it ignores load fluctuations and subtle differences between different dates, and is easily affected by short-term abnormal data, resulting in the selection of similar days that are not accurate enough.

[0044] In an optional embodiment, basic processing can also be implemented in other ways, such as through the quantile method, that is, using the quantile distribution of historical load data as a criterion for screening similar days; in this method, the quantile of the electricity consumption data of each date is first calculated (such as the median, quartile, etc.), and then the data of the day to be predicted is compared with the quantile of the historical data, and the date with the closest quantile is selected as the similar day; this method can avoid the influence of outliers to a certain extent and is suitable for relatively stable load patterns; however, the quantile method cannot accurately capture the trends and dynamic changes of electricity consumption patterns, and has poor adaptability to changes in seasonal or long-term trends, making it difficult to achieve accurate screening of similar days in a complex and changeable electricity demand environment.

[0045] Example 2, reference Figure 2-Figure 3 , is an embodiment of the present invention, which provides a data-driven power similar day screening method, including:

[0046] S2: Build the first model and train it to learn the first weights.

[0047] It should be noted that if Figure 1 S2 and Figure 2As shown, constructing the first model and training it includes constructing the first model using a neural network, learning the first weight distribution through a training process, and improving the accuracy of screening.

[0048] Furthermore, the first weight includes a similarity weight, which is used to measure the similarity of electricity consumption patterns between different historical dates and the day to be predicted; the similarity weight is obtained through a comprehensive calculation of the Kalman filter principle and the neural network model, and is corrected in the subsequent processing stage to ensure the accuracy of the screened similar days.

[0049] Furthermore, according to Kalman filter theory, the minimum variance estimation of power consumption pattern can be achieved by weighted fusion of multiple estimates:

[0050]

[0051] The weight λ i There is a clear analytical expression:

[0052]

[0053] In the formula, is the variance of the ith estimate. Obviously, when the variance is smaller, the corresponding weight will be larger, which also means that it is more similar to the actual power consumption pattern, which provides a basis for realizing similar day screening.

[0054] S3: Process the first weight.

[0055] It should be noted that if Figure 1 S3 and Figure 2 As shown, the processing of the first weight includes using a class of methods to preliminarily calculate the first weight and make corrections to ensure the accuracy of the first weight.

[0056] Furthermore, the first weight processing also includes performing sliding average and linear regression correction to eliminate irrelevant components.

[0057] Furthermore, although formula (5) gives the definition of weight, it cannot be solved directly because the variance of each estimate is unknown. However, the numerator and denominator have the same dimensions, and the weight is actually a dimensionless value, which means that it can be estimated in a data-driven way.

[0058] Learning weight distribution using data-driven methods requires determining the loss function, combining equations (1) and (4), namely:

[0059]

[0060] S4: Make a prediction based on the processed first weight.

[0061] It should be noted that if Figure 1 Medium S4, Figure 2 as well as Figure 3 As shown, making a prediction includes using the modified first weight to sort the historical data, selecting the data, and achieving accurate prediction.

[0062] Furthermore, based on S3, it means that the load of any day to be predicted can be expressed as a linear weighted input of historical data. The weight consists of two parts. The optimal weight estimation can minimize the prediction error ε′ t , so two neural networks can be trained to predict p′ respectively i and p″ i , the product of the two is used as the weight to weight the input historical data, and the loss function is minimized as the prediction error, that is:

[0063]

[0064] In the formula, M is the total number of samples involved in training, and N is the total number of historical dates involved in similar day screening. i and p″ i , this is because p′ i =λ i , there is ∑λ i =1 and λ>0; and p″ i =L T (t) / L T (i) represents the trend growth, and its value should be around 1; considering the above a priori design of the neural network is conducive to training convergence;

[0065] Consider using convolutional neural networks to fit the weight distribution, because the design of convolutional neural networks sharing weights can easily process high-dimensional data; specifically, 384 days of historical data before the day to be predicted are used as input, and two convolutional neural networks with exactly the same structure but no shared parameters are used to fit p′ respectively. i and p″ i ; For fitting p′ i The network, reference λ i , a softmax operation is performed before the output to ensure that the output is greater than zero and the sum is 1; and for fitting p″ i The network refers to L T (t) / L T (i) A bias of +1.0 is added before the output.

[0066] The above is a schematic scheme of a data-driven power similar day screening method of this embodiment. It should be noted that the technical scheme of the system of this data-driven power similar day screening method and the technical scheme of the above-mentioned data-driven power similar day screening method belong to the same concept. The details of the technical scheme of the data-driven power similar day screening system in this embodiment that are not described in detail can all be referred to the description of the technical scheme of the above-mentioned data-driven power similar day screening method.

[0067] Example 3, reference Figure 4-Figure 5 ,An embodiment of the present invention provides a data-driven electricity similar day screening method. ,In order to verify the beneficial effects of the present invention, scientific ,demonstration is carried out through economic benefit calculation and simulation ,experiments.

[0068] The double convolutional neural network can obtain the weight p i is estimated, but its component p′ i =λ i It directly represents the similarity, so it is necessary to eliminate another component p″ i =L T (t) / L T (i) to obtain the component p′ i . Although the neural network above fits the weight p″ i However, this patent uses sliding average and linear regression to estimate it more accurately. The steps are as follows Figure 4 shown.

[0069] First, the historical data is averaged over a 365-day window to obtain the growth trend of electricity consumption. Then, a linear regression is performed to remove short-term mutations caused by factors such as the Spring Festival to obtain the fitting curve y=ax+b (the red dotted line in the figure). Component p″ i =L T (t) / L T (i) can be calculated using (at+b) / (ai+b). i Divide by the corresponding (at+b) / (ai+b) to get the component p′ i =λ i This component directly represents the similarity. The larger it is, the stronger the similarity between the corresponding historical data and the day to be predicted. Therefore, we only need to sort them from large to small to get any number of similar days.

[0070] like Figure 5 As shown, for example, the input of the neural network is the historical data of the previous 384 days. The network will evaluate the weight of each historical data input (in the lower part of the figure, there are 384 weight values ​​for 384 days of input, displayed in a bar chart), and weight similar days according to the obtained weights to obtain the load forecast result (the black line in the upper part of the figure).

[0071] Generally speaking, the larger the weight, the higher the similarity. The upper half of the figure shows the historical data of the top 10 weights, and it can be seen that it is very close to the true label.

[0072] However, from formula (6), we know that the weights estimated by the neural network include trend terms, so we use the sliding average and linear regression method to correct the weights. Each weight is divided by the corresponding value. The corrected weights can more accurately represent the similarity. Sorting by the corrected weights can more accurately filter similar days.

[0073] Embodiment 4 is an embodiment of the present invention, which provides a data-driven power similar day screening system, including: a preliminary screening module, a model training module, a weight processing module and a prediction execution module;

[0074] The preliminary screening module uses first principles to make estimates and establish a preliminary screening basis;

[0075] The model training module constructs and trains a first model to learn a first weight;

[0076] The weight processing module processes the first weight;

[0077] The prediction execution module performs prediction based on the processed first weight.

[0078] This embodiment further provides a computing device, which is applicable to a data-driven power similar day screening method, and includes:

[0079] A memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute computer executable instructions to implement a data-driven power similar day screening method as proposed in the above embodiment.

[0080] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, a data-driven power similar day screening method as proposed in the above embodiment is implemented.

[0081] The storage medium proposed in this embodiment and the data-driven power similar day screening method proposed in the above embodiment belong to the same inventive concept. The technical details not fully described in this embodiment can be referred to the above embodiment, and this embodiment has the same beneficial effects as the above embodiment.

[0082] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc., which can store program codes.

[0083] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute instructions), or in conjunction with such instruction execution systems, devices or apparatuses. For the purposes of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in conjunction with such instruction execution systems, devices or apparatuses.

[0084] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, a plurality of steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0085] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A data-driven power similar day screening method, characterized by: include: Use first principles to make estimates and establish a preliminary screening basis; Build the first model and train it to learn the first weights; Processing the first weight; Based on the processed first weight, prediction is made.

2. A data-driven power similar day screening method as claimed in claim 1, characterized in that: The establishing of the preliminary screening basis includes making a preliminary estimate of the historical data based on first principles to provide a basic basis for subsequent screening.

3. A data-driven power similar day screening method as claimed in claim 2, characterized in that: The constructing and training of the first model includes constructing the first model using a neural network, learning the first weight distribution through a training process, and improving the accuracy of screening.

4. A data-driven power similar day screening method as claimed in claim 3, characterized in that: The processing of the first weight includes using a class of methods to preliminarily calculate the first weight and making corrections to ensure the accuracy of the first weight.

5. A data-driven power similar day screening method as claimed in claim 4, characterized in that: The prediction includes sorting the historical data using the modified first weight and selecting the data, thereby laying a foundation for accurate prediction.

6. A data-driven power similar day screening method as claimed in claim 5, characterized in that: The first principle includes the Kalman filter principle, which estimates the similarity of power consumption patterns on different dates through basic processing of historical load data, providing a reliable basis for screening similar days; Kalman filtering can perform minimum variance estimation on historical data and obtain the first weight of the actual power consumption pattern.

7. A data-driven power similar day screening method as claimed in claim 6, characterized in that: The first weight includes a similarity weight, which is used to measure the similarity of the electricity consumption patterns between different historical dates and the day to be predicted; The similarity weight is obtained through the comprehensive calculation of Kalman filter principle and neural network model, and is corrected in the subsequent processing stage to ensure the accuracy of the selected similar days; The processing of the first weight also includes performing sliding average and linear regression correction to eliminate irrelevant components.

8. A data-driven power similar day screening system based on any one of claims 1 to 7, characterized in that: include: Preliminary screening module, model training module, weight processing module and prediction execution module; The preliminary screening module uses first principles to make estimates and establish a preliminary screening basis; The model training module constructs and trains a first model to learn a first weight; The weight processing module processes the first weight; The prediction execution module performs prediction based on the processed first weight.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of a data-driven power similar day screening method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of a data-driven power similar day screening method according to any one of claims 1 to 7 are implemented.

Images