Method and system for generating provincial power grid load time sequence scenarios under critical weather conditions for supply guarantee

Abstract

The present invention relates to a method and system for generating a provincial power grid load time sequence scenario under critical weather conditions for supply guarantee, and belongs to the field of power system engineering technology. The steps are as follows: (1) proposing a basic load and a meteorological sensitive load, and dividing the meteorological sensitive load into a meteorological sensitive load utilization rate and a meteorological sensitive load peak capacity; (2) using a gray system model to estimate the basic load and meteorological sensitive load peak capacity of the future target year; (3) generating a meteorological sensitive load utilization rate for the future target year; (4) multiplying the meteorological sensitive load utilization rate of step (3) by the meteorological sensitive load peak capacity of step (2) to obtain a generated meteorological sensitive load, and the generated meteorological sensitive load plus the basic load of step (2) is the generated provincial load power, that is, the provincial power grid load time sequence scenario under critical weather processes for supply guarantee. The present invention improves the correlation between power load and factors such as social and economic development, population consumption level, and meteorological processes.

Description

Technical Field

[0001] The present invention relates to a method and system for generating a provincial power grid load time sequence scenario under critical weather conditions for ensuring power supply, and belongs to the technical field of power system engineering. Background Art

[0002] Affected by meteorological changes, the impact of meteorology on load power in new power systems with a high proportion of renewable energy access in the future will gradually increase, and the volatility and uncertainty of power load will become more and more obvious. On the one hand, the national economy continues to develop with high quality, and the cold and hot loads that are greatly affected by meteorology such as temperature and humidity are increasing; on the other hand, global climate change has intensified the frequency of extreme weather, especially the continuous extreme heat and windless, extreme cold and lightless key meteorological events under cold waves and heat waves, which will aggravate the contradiction of power balance in new power systems. The scenario generation method is one of the effective methods to quantify load uncertainty, which can be used to construct random scenarios that meet the time series characteristics. Therefore, constructing a load scenario generation model for key weather conditions for power supply can provide important support for the safe and stable operation of the power system, economic dispatch, and reasonable planning of the power grid.

[0003] At present, the application of scenario generation methods in power load is mainly divided into probability models, traditional time series models and deep learning models. The basic idea of ​​the scenario generation method based on probability models is to assume that random fluctuations such as loads satisfy a certain probability distribution such as normal distribution, and then obtain the simulation scenario through sampling methods. This method is widely used in scenario generation methods based on Monte Carlo simulation, Gaussian process, and Latin hypercube sampling because of its simplicity and practicality. However, due to excessive simplifications and assumptions, the generated scenarios may deviate greatly from the actual scenarios. The scenario generation technology based on time series methods is another widely used method. Some researchers have combined the autoregressive conditional heteroskedastic model to model the time series characteristics of noise. However, the time series model is only suitable for fitting the linear combination relationship between random variables and noise, and cannot describe complex nonlinear relationships. Thanks to the advantage of deep neural networks that can effectively fit high-dimensional nonlinear mapping relationships, the scenario generation method based on deep learning models has developed rapidly in recent years. Some researchers have modeled the time series curves on a monthly scale based on multi-layer perceptrons to predict the load size in the future power system. Some researchers have used load key values ​​and similar curve data sets as conditions and training sets to build a load curve generation model based on CGAN. Some have also used manually defined meteorological label data to guide the generation of corresponding source-load scenarios.

[0004] The above-mentioned scenarios and methods for generating power loads have enriched the technology for time series scenarios under load uncertainty, and provided a reference for further studying the balance arrangement of power and electricity under new power systems. However, most of the current studies on power load scenario generation are based on typical day sets as the target object, and large-scale historical scenarios are reduced by using data mining technology and scenario reduction technology to generate a set of classic scenarios that can represent the entire problem interval. In the context of high proportion of renewable energy access, the risk of extreme scenarios is amplified. Therefore, it is necessary to pay attention to the generation of high-load scenarios that are difficult to ensure supply under critical weather conditions of cold waves and heat waves. Moreover, most previous studies did not consider the impact of meteorological factors on load power scenarios during model training, or directly constructed the mapping relationship between the two as input and output, lacking the understanding of the use of meteorological sensitive load components. Therefore, the idea of ​​load decomposition is used to explore the basic load components affected by macroeconomic development and the meteorological sensitive components that are greatly affected by meteorological factors such as temperature, and the load scenarios under critical weather conditions of cold waves and heat waves in the future target year are obtained by superposition, providing reference information for the power grid dispatching department to optimize the allocation of power grid resources. Summary of the invention

[0005] In view of the deficiencies in the prior art, the present invention provides a method and system for generating a provincial power grid load timing scenario under critical weather conditions for ensuring power supply.

[0006] The technical solution of the present invention is as follows:

[0007] A method for generating a provincial power grid load time sequence scenario under critical weather conditions for supply guarantee, the steps are as follows:

[0008] (1) Based on the provincial load power and historical meteorological data of the load center, the basic load and meteorological sensitive load are proposed, and the meteorological sensitive load is divided into two parts: the meteorological sensitive load utilization rate and the meteorological sensitive load peak capacity;

[0009] (2) Use the grey system model to estimate the base load and peak capacity of meteorological sensitive load in the future target year;

[0010] (3) Combine the CNN convolutional neural network that reflects spatial information and the LSTM model that can extract key information of long time series to generate the meteorological sensitive load utilization rate in the future target year;

[0011] (4) Multiply the meteorological sensitive load utilization rate obtained in step (3) by the meteorological sensitive load peak capacity obtained in step (2) to obtain the generated meteorological sensitive load. The generated meteorological sensitive load plus the basic load obtained in step (2) is the generated provincial load power, that is, the provincial power grid load timing scenario under the key weather process for supply guarantee.

[0012] Preferably, in step (1) of the present invention, the specific steps are as follows:

[0013] (1-1) Obtain provincial load power (which can be obtained through the power grid dispatching department) and historical meteorological data of the load center (which can be obtained through meteorological websites, including ambient temperature, relative humidity, wind speed, etc.);

[0014] (1-2) Input meteorological elements such as temperature and humidity to calculate the effective body temperature, and select dates in spring and autumn that meet the meteorological range of suitable effective body temperature as spring base load days and autumn base load days;

[0015] (1-3) Since the correlation between the meteorological sensitive load calculated under the daily load on different base load days and various meteorological factors is different, the correlation weights of different base load days are considered to establish the summer base load and winter base load respectively;

[0016] (1-4) Subtract the provincial load obtained in step (1) from the summer basic load and the winter basic load to obtain the summer meteorological sensitive load and the winter meteorological sensitive load. The meteorological sensitive load is further divided into two parts for analysis: the meteorological sensitive load utilization rate under the influence of meteorological changes and the meteorological sensitive load peak capacity considering non-meteorological factors such as economic and cultural factors.

[0017] Preferably, in step (1-1), a conditional generative adversarial network is used to perform data enhancement on the acquired provincial load power and load center meteorological historical data, and the enhancement steps are:

[0018] For the enhancement of load sample data under cold and heat waves, the peak-to-valley difference rate, load rate, skewness, and kurtosis are used as conditional variables, and the random sequence obtained by random sampling is used as the noise input. Different extreme load indexes are added to guide the generation of load samples under corresponding key weather conditions.

[0019] For matching meteorological sample data enhancement, the conditional variables are month labels and meteorological categories, and the noise sequence obtained by random sampling is used as input to guide the generation of meteorological sample data under corresponding key weather conditions.

[0020] According to the preferred embodiment of the present invention, in step (1-2), the effective body temperature T f The calculation method is:

[0021]

[0022] In the formula, T m is the outdoor temperature in Celsius (℃); R h is the relative humidity at the moment (%); V is the wind speed (m / s);

[0023] When the effective perceived temperature is between 16-25℃, it is considered that there is no meteorological sensitive load. All dates in spring April to May and autumn September to October that meet the above-defined temperature range of 16-25℃ are selected as spring basic load days and autumn basic load days.

[0024] According to the preferred embodiment of the present invention, in steps (1-3), the summer base load and the winter base load are:

[0025]

[0026] In the formula, P Bx and P Bd They are summer base load and winter base load respectively; P s,ij and P a,ij Spring s Month and Autumn a Month i Tiandi j Load value at the moment; W s,i and W a,i are the correlation weights for spring base load days and autumn base load days, respectively;

[0027] The calculation method for the correlation weights of spring base load days and autumn base load days is:

[0028]

[0029] In the formula, J x,i and J d,i Select spring i Base load day and fall i The Spearman correlation coefficient between the meteorological sensitive load and the corresponding effective body temperature in summer (June, July, and August) and winter (January, February, and December) calculated on the base load day; J x, and J d, The numerical sum of the above correlation coefficients was calculated for all base load days in spring and autumn respectively; m , n The total number of days with a temperature between 16 and 25°C in spring and autumn respectively.

[0030] Preferably, in step (1-4) of the present invention, the calculation formulas for the summer meteorological sensitive load and the winter meteorological sensitive load are:

[0031]

[0032] In the formula, P L is the provincial load power, P B is the summer base load or winter base load, represents random volatility load, which can be ignored due to its small disturbance effect;

[0033] The meteorological sensitive load utilization rate and meteorological sensitive load peak capacity are expressed as:

[0034]

[0035]

[0036] In the formula, P weather is the meteorological sensitive load component, subscript t Indicates the daily load curve t Meteorological sensitive load value at each moment; C e The peak capacity of meteorological sensitive load. Due to the limitation of collected data, the peak capacity of meteorological sensitive load needs to be estimated. The estimation method is the maximum value of all meteorological sensitive load values ​​in that year. λ It is the utilization rate of meteorological sensitive loads, which reflects the overall utilization of equipment such as cooling (heating) that is highly correlated with the weather.

[0037] Preferably, according to the present invention, in step (2), the grey system model is:

[0038]

[0039] In the formula, is the original characteristic sequence, i.e. the basic load value sequence at the same time of each historical year and quarter; is the original feature sequence The first-order AGO accumulation sequence of ; is the sequence of influencing factors, i.e. the sequence of national economic indicators corresponding to the historical years (economic indicators include gross national product (GDP), consumer price index (CPI), added value of real estate industry, proportion of tertiary industry and residents' electricity consumption, etc.), subscript i Indicates i economic and humanistic indicators, i =1,2,…, N , N is the total number of indicators, for The first-order AGO accumulation sequence of ; k is the time value, k=2,3,…,96; a , b , c , d are model parameters, estimated by the least squares method; η is the background value coefficient, and the optimal value of the background value coefficient is optimized by the particle swarm optimization algorithm;

[0040] After the model training is completed, given the historical national economic indicators and historical quarterly basic loads, the grey system model is used to estimate the basic load of the future target year and the peak capacity of the meteorological sensitive load (when estimating the peak capacity of the meteorological sensitive load, the original feature sequence can be modified to the peak capacity of the meteorological sensitive load of the historical year).

[0041] Preferably, in step (3), the model includes an input layer, a convolutional layer, a flattening layer, an LSTM layer, a fully connected layer and an output layer.

[0042] The input layer is a time series matrix of multiple meteorological elements distribution , Represents the current t The provincial meteorological element data for the day, each matrix The dimension is , m is the number of cities, n is the time resolution, s is the type of meteorological element, D Indicates the number of historical days to look ahead, which can be determined according to the classification results of key weather processes. For example, the number of days to look ahead for a large-scale heat wave / cold wave is 3 days, and the number of days to look ahead for a small- to medium-scale heat wave / cold wave is 1 day;

[0043] The convolution calculation of the convolution layer is as follows:

[0044]

[0045] in, For the neural network i Tier j The feature samples are located in The value at They are the three directional dimensions of the convolution kernel; g is the activation function of the neuron; m is the number of connection features of the network in the previous stage; Indicates i Layer j Position in feature map The convolution kernel is connected to the Tier m The weight of the feature map; For the i Layerj The bias of the feature;

[0046] The LSTM layer controls the timing state of information flow in the network through three gate logic units: input gate, forget gate, and output gate. This enables it to automatically and selectively forget and retain relevant feature information during the time series feature learning process, thereby effectively avoiding the occurrence of gradient vanishing or gradient exploding problems.

[0047] Forget Gate Used to determine the internal state result at the previous moment The calculation formula for what information needs to be discarded at the current moment is as follows:

[0048]

[0049] In the formula, Indicates the output result at the last moment; Indicates the current moment input; W f and b f is the forget gate weight matrix and bias parameter; Represents the sigmoid activation function;

[0050] Input Gate Determine the calculation result of the candidate state at the current moment Which information needs to be saved? The calculation formula is:

[0051]

[0052]

[0053] In the formula, W c and b c Calculate the candidate states at the current moment respectively The weight matrix and bias parameters used; tanh (•) indicates the corresponding activation function; W i and b i Represents the input gate weight matrix and bias parameters;

[0054] Based on the forget gate and input gate, the internal state at the current moment can be calculated , and its calculation formula is as follows:

[0055]

[0056] Based on the calculation result of the internal state at the current moment, the result obtained under the activation function is combined with the output gate Perform Hadamard product operation to get the final output at the current moment , the calculation formula is as follows:

[0057]

[0058]

[0059] In the formula, W o and b o Represent the output gate weight matrix and bias parameters respectively;

[0060] Finally, the output results of extracting temporal characteristics and spatial correlation are obtained, and together with the economic indicators for evaluating electricity consumption habits, they are used as the input variables of the fully connected layer. The output result of the fully connected layer is the utilization rate of meteorological sensitive loads in the corresponding period.

[0061] A provincial power grid load time sequence scenario generation system under critical weather conditions for supply guarantee, comprising:

[0062] Analysis module: It is used to propose basic load and meteorological sensitive load by integrating provincial load power and historical meteorological data of load center, and divide meteorological sensitive load into two parts: meteorological sensitive load utilization rate and meteorological sensitive load peak capacity;

[0063] Grey system model estimation module: used to estimate the base load and meteorological sensitive load peak capacity in the future target year using the grey system model;

[0064] Deep learning model estimation module: used to combine the CNN convolutional neural network that reflects spatial information and the LSTM model that can extract key information of long time series to generate the meteorological sensitive load utilization rate in the future target year;

[0065] Generation module: used to multiply the utilization rate of the meteorological sensitive load in the future target year by the peak capacity of the meteorological sensitive load in the future target year to obtain the generated meteorological sensitive load. The generated meteorological sensitive load plus the basic load in the future target year is the generated provincial load power, that is, to generate the provincial power grid load timing scenario under the key weather process for supply guarantee.

[0066] The beneficial effects of the present invention are:

[0067] Different from directly using traditional statistical probability models or time series extrapolation analysis, the present invention establishes the mapping relationship between the target year base load and the meteorological sensitive load based on the idea of ​​load decomposition, improves the correlation between power load and social and economic development, population consumption level, meteorological process and other factors, and uses a deep learning model with the ability to learn high-dimensional nonlinear modeling after sample data enhancement. The classification modeling of key weather processes that affect power supply is more conducive to the precise grasp of the scenario simulation of meteorological sensitive load utilization rate under weather processes of different severity of power load, and superimposed generation of load time series under the action of key weather processes. Therefore, the present invention is an effective scenario analysis method that can reflect the changes in load time series under the action of extreme meteorological processes. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 This is a schematic diagram of the process of Example 1 of the present invention;

[0069] Figure 2 This is a schematic diagram of calculating the daily correlation weights of different base loads in Example 1 of the present invention;

[0070] Figure 3 This is a flow chart of a grey system model according to Embodiment 1 of the present invention;

[0071] Figure 4 A conditional generation adversarial network game training framework according to Embodiment 1 of the present invention;

[0072] Figure 5 This is a schematic diagram of a scenario simulation of utilization rate of classified meteorological sensitive loads based on CNN-LSTM according to Embodiment 1 of the present invention;

[0073] Figure 6 This is a schematic diagram of scene generation results under a heat wave critical weather process according to Example 1 of the present invention;

[0074] Figure 7 This is a schematic diagram of scene generation results under the critical weather process of cold wave in Example 1 of the present invention. DETAILED DESCRIPTION

[0075] The present invention will be further described below by way of embodiments in conjunction with the accompanying drawings, but is not limited thereto.

[0076] Embodiment 1:

[0077] like Figure 1 As shown, this embodiment provides a method for generating a provincial power grid load time sequence scenario under critical weather conditions for supply guarantee, and the steps are as follows:

[0078] (1) Based on the provincial load power and the historical meteorological data of the load center, the basic load and meteorological sensitive load are proposed, and the meteorological sensitive load is divided into two parts: the meteorological sensitive load utilization rate and the meteorological sensitive load peak capacity. The specific steps are as follows:

[0079] (1-1) Obtain provincial load power (which can be obtained through the power grid dispatching department) and historical meteorological data of the load center (which can be obtained through meteorological websites, including ambient temperature, relative humidity, wind speed, etc.);

[0080] The conditional generative adversarial network is used to enhance the data of provincial load power and load center meteorological historical data. It combines supervised learning technology with generative adversarial networks, and the network training is more targeted. It obtains generated samples that are as similar as possible to real data through adversarial training of generator G and discriminator D. The task of generator G is to generate sample scenarios that meet the conditions, and the task of the discriminator is to judge the similarity between the generated scenarios and the actual ones, output the discriminant value and pass it back to itself and generator G, so that the two can continuously compete and optimize their own capabilities until the confrontation reaches a balance.

[0081] The loss functions of the generator and discriminator can be expressed as:

[0082]

[0083] In the formula, and They represent the corresponding real samples x and random noise z The expectation of the distribution satisfied; Indicates that the condition c Next Generator G Output generated samples; D (.) represents the discriminator score function.

[0084] The total target loss function of the conditional generative adversarial network after Wasserstein distance optimization is:

[0085]

[0086] In the formula, s It is the variable after interpolation of generated samples and real data; Denotes the discriminator function The second norm of the gradient matrix; Represents the gradient penalty coefficient.

[0087] In terms of load sample data enhancement under cold and heat waves, the peak-to-valley difference rate, load rate, skewness, and kurtosis are used as conditional variables, and the random sequence obtained by random sampling is used as noise as input. Different extreme load index metrics are added to guide the generation of load samples under corresponding key weather conditions. For matching meteorological sample data enhancement, the conditional variables are the month label and meteorological category, and the noise sequence obtained by random sampling is used as input to guide the generation of meteorological sample data under corresponding key weather conditions.

[0088] The above conditions generate the adversarial network game training framework process, which can be found in the attached Figure 4 .

[0089] (1-2) Input meteorological elements such as temperature and humidity to calculate the effective body temperature, and select dates in spring and autumn that meet the meteorological range of suitable effective body temperature as spring base load days and autumn base load days;

[0090] Effective body temperature T f The calculation method is:

[0091]

[0092] In the formula, T m is the outdoor temperature in Celsius (℃); R h is the relative humidity at the moment (%); V is the wind speed (m / s);

[0093] When the effective perceived temperature is between 16-25℃, it is considered that there is no meteorological sensitive load. All dates in spring April to May and autumn September to October that meet the above-defined temperature range of 16-25℃ are selected as spring basic load days and autumn basic load days.

[0094] (1-3) Since the correlation between the meteorological sensitive load calculated under the daily load on different base load days and various meteorological factors is different, the correlation weights of different base load days are considered to establish the summer base load and winter base load respectively;

[0095] The summer base load and winter base load are:

[0096]

[0097] In the formula, P Bx and P Bd They are summer base load and winter base load respectively; P s,ij and P a,ij Spring s Month and Autumn a Month i Tiandi j Load value at the moment; W s,i and W a,i are the correlation weights for spring base load days and autumn base load days, respectively;

[0098] The calculation method for the correlation weights of spring base load days and autumn base load days is:

[0099]

[0100] In the formula, J x,i and J d,i Select spring i Base load day and fall i The Spearman correlation coefficient between the meteorological sensitive load and the corresponding effective body temperature in summer (June, July, and August) and winter (January, February, and December) calculated on the base load day; J x, and J d, The numerical sum of the above correlation coefficients was calculated for all base load days in spring and autumn respectively; m , n The total number of days with a temperature between 16 and 25°C in spring and autumn respectively.

[0101] (1-4) Subtract the provincial load obtained in step (1) from the summer basic load and the winter basic load to obtain the summer meteorological sensitive load and the winter meteorological sensitive load. The meteorological sensitive load is further divided into two parts for analysis: the meteorological sensitive load utilization rate under the influence of meteorological changes and the meteorological sensitive load peak capacity considering non-meteorological factors such as economic and cultural factors.

[0102] The calculation formulas for summer meteorological sensitive load and winter meteorological sensitive load are as follows:

[0103]

[0104] In the formula, P L is the provincial load power, P B is the summer base load or winter base load, represents random volatility load, which can be ignored due to its small disturbance effect;

[0105] The meteorological sensitive load utilization rate and meteorological sensitive load peak capacity are expressed as:

[0106]

[0107]

[0108] In the formula, P weather is the meteorological sensitive load component, subscript t Indicates the daily load curve tMeteorological sensitive load value at each moment; C e The peak capacity of meteorological sensitive load. Due to the limitation of collected data, the peak capacity of meteorological sensitive load needs to be estimated. The estimation method is the maximum value of all meteorological sensitive load values ​​in that year. λ It is the utilization rate of meteorological sensitive loads, which reflects the overall utilization of equipment such as cooling (heating) that is highly correlated with the weather.

[0109] (2) The grey system model is used to estimate the base load and peak capacity of meteorological sensitive load in the future target year. Compared with artificial intelligence methods that require a large amount of data for training, the grey system model has the advantage of being less dependent on the amount of data and is suitable for fitting and simulating known data sequences under the premise of a small amount of data. In order to consider the influence of multiple factors with the same base load or correlation, an improved multivariable grey system model is used for simulation.

[0110] The grey system model is:

[0111]

[0112] In the formula, is the original characteristic sequence, i.e. the basic load value sequence at the same time of each historical year and quarter; is the original feature sequence The first-order AGO accumulation sequence of ;

[0113]

[0114] is the sequence of influencing factors, i.e. the sequence of national economic indicators corresponding to the historical years (economic indicators include gross national product (GDP), consumer price index (CPI), added value of real estate industry, proportion of tertiary industry and residents' electricity consumption, etc.), subscript i Indicates i economic and humanistic indicators, i =1,2,…, N , N is the total number of indicators;

[0115] for The first-order AGO accumulation sequence is:

[0116]

[0117] In the formula, k is the time value, k =2,3,…,96; a , b , c , dare model parameters, estimated by the least squares method, is the parameter matrix, , the specific expression is:

[0118]

[0119]

[0120] η is the background value coefficient, and the optimal value of the background value coefficient is optimized by the particle swarm optimization algorithm;

[0121] The simulated value of the original feature sequence is obtained by cumulative reduction, as follows:

[0122]

[0123] In the formula, ; Represents the rth simulated value of the original feature sequence.

[0124] The particle swarm algorithm has the advantages of fast convergence speed, few parameters, simple algorithm and easy implementation. The algorithm process is as follows:

[0125] 1) Initialize all parameters of the population particles, as well as the positions and velocities of the particles;

[0126] 2) Calculate the fitness function value of each solution in the current population. The fitness function is:

[0127]

[0128] 3) Find the individual historical optimal position of the particle and the global historical optimal position of the particle group;

[0129] 4) Update the particle position and parameters according to the inertia weight, its own speed, and the distance and direction difference from the optimal position of the historical individual itself and the particle group;

[0130] 5) If the termination condition and the number of iterations are not met, repeat the above steps 2) to 4); if they are met, output the optimal value of the background value coefficient η ;

[0131] The process of the above grey system model is shown in the attached Figure 3 .

[0132] After the model training is completed, given the national economic indicators of the historical year and the historical quarterly basic load, the grey system model is used to estimate the basic load of the future target year and the peak capacity of the meteorological sensitive load.

[0133] (3) Combine the CNN convolutional neural network that reflects spatial information and the LSTM model that can extract key information of long time series to generate the meteorological sensitive load utilization rate in the future target year;

[0134] The model includes input layer, convolution layer, flattening layer, LSTM layer, fully connected layer and output layer.

[0135] The input layer is a time series matrix of multiple meteorological elements distribution , Represents the current t The provincial meteorological element data for the day, each matrix The dimension is , m is the number of cities, n is the time resolution, s is the type of meteorological element, D Indicates the number of historical days to look ahead, which can be determined according to the classification results of key weather processes. For example, the number of days to look ahead for a large-scale heat wave / cold wave is 3 days, and the number of days to look ahead for a small- to medium-scale heat wave / cold wave is 1 day;

[0136] The convolution calculation of the convolution layer is as follows:

[0137]

[0138] in, For the neural network i Tier j The feature samples are located in The value at They are the three directional dimensions of the convolution kernel; g is the activation function of the neuron; m is the number of connection features of the network in the previous stage; Indicates i Layer j Position in feature map The convolution kernel is connected to the Tier m The weight of the feature map; For the i Layer j The bias of the feature;

[0139] When building network training for the meteorological sensitive load utilization scenario, the input is a data tensor matrix composed of historical meteorological data. The convolution kernel slides in the horizontal and vertical dimensions, which can effectively learn the spatial characteristics of the input data. The output is the abstract features extracted by convolution and expanded as the input of the LSTM layer.

[0140] The LSTM layer controls the timing state of information flow in the network through three gate logic units: input gate, forget gate, and output gate. This enables it to automatically and selectively forget and retain relevant feature information during the time series feature learning process, thereby effectively avoiding the occurrence of gradient vanishing or gradient exploding problems.

[0141] Forget Gate Used to determine the internal state result at the previous moment The calculation formula for what information needs to be discarded at the current moment is as follows:

[0142]

[0143] In the formula, Indicates the output result at the last moment; Indicates the current moment input; W f and b f is the forget gate weight matrix and bias parameter; Represents the sigmoid activation function;

[0144] Input Gate Determine the calculation result of the candidate state at the current moment Which information needs to be saved? The calculation formula is:

[0145]

[0146]

[0147] In the formula, W c and b c Calculate the candidate states at the current moment respectively The weight matrix and bias parameters used; tanh (•) indicates the corresponding activation function; W i and b i Represents the input gate weight matrix and bias parameters;

[0148] Based on the forget gate and input gate, the internal state at the current moment can be calculated , and its calculation formula is as follows:

[0149]

[0150] Based on the calculation result of the internal state at the current moment, the result obtained under the activation function is combined with the output gate Perform Hadamard product operation to get the final output at the current moment , the calculation formula is as follows:

[0151]

[0152]

[0153] In the formula, W o and b o Represent the output gate weight matrix and bias parameters respectively;

[0154] Finally, the output results of extracting temporal characteristics and spatial correlation are obtained, and together with the economic indicators for evaluating electricity consumption habits, they are used as the input variables of the fully connected layer. The output result of the fully connected layer is the utilization rate of meteorological sensitive loads in the corresponding period.

[0155] (4) Multiply the meteorological sensitive load utilization rate obtained in step (3) by the meteorological sensitive load peak capacity obtained in step (2) to obtain the generated meteorological sensitive load. The generated meteorological sensitive load plus the basic load obtained in step (2) is the generated provincial load power, that is, the provincial power grid load timing scenario under the key weather process for supply guarantee.

[0156] In order to verify the technical concept of this embodiment, a specific example analysis is given below.

[0157] Taking a certain province as the research object, we collected economic and cultural data such as the province's gross domestic product (GDP), consumer price index (CPI), real estate industry added value, tertiary industry and residents' electricity consumption from 2017 to 2021, as well as the corresponding load power data and historical data of meteorological elements in the 16 prefecture-level cities under its jurisdiction (the resolution of load and meteorological data is 15 minutes).

[0158] Taking 2021 as an example, the results of load scenario generation under key weather conditions of cold and heat waves obtained by the above method are as follows: Figure 6 and Figure 7 With a good judgment on the changes in the macro social environment, this example can more accurately reflect the changing trend of loads under key weather processes and construct time series scenarios.

[0159] Embodiment 2:

[0160] This embodiment provides a provincial power grid load time sequence scenario generation system under critical weather conditions for ensuring power supply, including:

[0161] Analysis module: It is used to propose basic load and meteorological sensitive load by integrating provincial load power and historical meteorological data of load center, and divide meteorological sensitive load into two parts: meteorological sensitive load utilization rate and meteorological sensitive load peak capacity;

[0162] Grey system model estimation module: used to estimate the base load and meteorological sensitive load peak capacity in the future target year using the grey system model;

[0163] Deep learning model estimation module: used to combine the CNN convolutional neural network that reflects spatial information and the LSTM model that can extract key information of long time series to generate the meteorological sensitive load utilization rate in the future target year;

[0164] Generation module: used to multiply the utilization rate of the meteorological sensitive load in the future target year by the peak capacity of the meteorological sensitive load in the future target year to obtain the generated meteorological sensitive load. The generated meteorological sensitive load plus the basic load in the future target year is the generated provincial load power, that is, to generate the provincial power grid load timing scenario under the key weather process for supply guarantee.

[0165] Those skilled in the art will appreciate that the units, i.e., algorithm steps, of the various examples described in conjunction with this embodiment can be implemented in electronic hardware or in a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0166] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A method for generating a provincial power grid load time sequence scenario under critical weather conditions, characterized in that: Here are the steps: (1) Based on the provincial load power and the historical meteorological data of the load center, the basic load and meteorological sensitive load are proposed, and the meteorological sensitive load is divided into two parts: the meteorological sensitive load utilization rate and the meteorological sensitive load peak capacity; The specific steps are as follows: (1-1) Obtain historical data on provincial load power and load center meteorology; (1-2) Calculate the effective body temperature and select dates in spring and autumn that meet the appropriate meteorological range for the effective body temperature as spring base load days and autumn base load days; Effective body temperature T f The calculation method is: Where, T m is the outdoor temperature in degrees Celsius; R h is the relative humidity at the moment; V is the wind speed; When the effective body temperature is between 16 and 25°C, it is considered that there is no meteorological sensitive load, and all dates in spring from April to May and autumn from September to October that meet the temperature range of 16 to 25°C are selected as spring basic load days and autumn basic load days; (1-3) Considering the daily correlation weights of different base loads, the summer base load and winter base load are established respectively; The summer base load and winter base load are: Where P Bx and P Bd are summer base load and winter base load respectively; P s,ij and P a,ij are the load values ​​at the jth moment on the i-th day in spring month s and autumn month a respectively; W s,i and W a,i are the correlation weights for spring base load days and autumn base load days, respectively; The calculation method for the correlation weights of spring base load days and autumn base load days is: In the formula, J x,i and J d,i are the Spearman correlation coefficients between the summer and winter meteorological sensitive loads and the corresponding effective body temperature calculated on the i-th base load day in spring and the i-th base load day in autumn, respectively; J x , and J d, The sum of the values ​​of the above correlation coefficients calculated for all base load days in spring and autumn respectively; m and n are the total number of days with a temperature between 16 and 25 °C in spring and autumn respectively; (1-4) Subtract the provincial load obtained in step (1) from the summer basic load and the winter basic load to obtain the summer meteorological sensitive load and the winter meteorological sensitive load, and further analyze the meteorological sensitive load into two parts: meteorological sensitive load utilization rate and meteorological sensitive load peak capacity; (2) Use the grey system model to estimate the base load and peak capacity of meteorological sensitive load in the future target year; (3) Combine the CNN convolutional neural network that reflects spatial information and the LSTM model that can extract key information of long time series to generate the meteorological sensitive load utilization rate in the future target year; (4) Multiplying the meteorological sensitive load utilization rate obtained in step (3) by the meteorological sensitive load peak capacity obtained in step (2) to obtain the generated meteorological sensitive load. The generated meteorological sensitive load plus the basic load obtained in step (2) is the generated provincial load power, that is, the provincial power grid load timing scenario under the key weather process for supply guarantee.

2. The method for generating a provincial power grid load time sequence scenario under critical weather conditions for supply guarantee as claimed in claim 1, characterized in that: In step (1-1), a conditional generative adversarial network is used to perform data enhancement on the acquired provincial load power and load center meteorological historical data.

3. The method for generating a provincial power grid load time sequence scenario under critical weather conditions for supply guarantee as claimed in claim 2, characterized in that: In steps (1-4), the calculation formulas for summer meteorological sensitive load and winter meteorological sensitive load are: P weather =P L -P B -ξ s Where P L is the provincial load power, P B is the summer base load or winter base load, ξ s represents random volatility load; The meteorological sensitive load utilization rate and meteorological sensitive load peak capacity are expressed as: C e =max(max(P weather,t |t=1,2,…,96)) Where P weather is the meteorological sensitive load component, and the subscript t represents the meteorological sensitive load value at the tth moment on the daily load curve; C e is the peak capacity of the meteorological sensitive load. Due to the limitation of collected data, the peak capacity of the meteorological sensitive load needs to be estimated. The estimation method is the maximum value of all meteorological sensitive load values ​​in that year; λ is the utilization rate of the meteorological sensitive load.

4. The method for generating a provincial power grid load time sequence scenario under critical weather conditions for supply guarantee as claimed in claim 3, characterized in that: In step (2), the grey system model is: In the formula, is the original characteristic sequence, i.e. the basic load value sequence at the same time of the historical year and quarter; x1 (1) is the original feature sequence The first-order AGO accumulation sequence of ; is the sequence of influencing factors, that is, the sequence of national economic indicators corresponding to historical years. The subscript i represents the i-th economic and cultural indicator, i = 1, 2, ..., N, N is the total number of indicators, for The first-order AGO accumulation sequence of the model is: k is the time value, k = 2, 3, ..., 96; a, b, c, d are model parameters, estimated by the least squares method; η is the background value coefficient; After the model training is completed, given the national economic indicators of the historical year and the historical quarterly basic load, the grey system model is used to estimate the basic load of the future target year and the peak capacity of the meteorological sensitive load.

5. The method for generating a provincial power grid load time sequence scenario under critical weather conditions for supply guarantee as claimed in claim 4, characterized in that: In step (3), the model includes input layer, convolution layer, flattening layer, LSTM layer, fully connected layer and output layer. The input layer is the time series matrix of multiple meteorological elements distribution X = (X t-D ,…,X t-1 ,X t ), X t Represents the meteorological element data of the whole province on the current day t. Each matrix X t The dimension is m×n×s, where m is the number of cities, n is the time resolution, s is the type of meteorological elements, and D represents the number of historical days ahead; The convolution calculation of the convolution layer is as follows: in, is the value of the jth feature sample at (x, y, z) in the i-th layer of the neural network; P i ,Q i ,R i are the three directional dimensions of the convolution kernel; g is the activation function of the neuron; m is the number of connection features of the network in the previous stage; b represents the weight of the convolution kernel at position (p, q, r) in the jth feature map of the i-th layer connected to the mth feature map of the i-1th layer; ij is the bias of the jth feature of the i-th layer; The LSTM layer controls the timing state of information flow in the network through three gate logic units: input gate, forget gate, and output gate; Forget Gate t Used to determine the internal state result C at the previous moment t-1 The calculation formula for what information needs to be discarded at the current moment is as follows: f t =σ(W f ·[h t-1 ,x t ]+b f ) In the formula, h t-1 Indicates the output result at the last moment; x t Indicates the current input; W f and b f is the forget gate weight matrix and bias parameter; σ(g) represents the sigmoid activation function; Input Gate i t Determine the calculation result of the candidate state at the current moment Which information needs to be saved? The calculation formula is: Where W c and b c Calculate the candidate states at the current moment respectively The weight matrix and bias parameters used; tanh(·) represents the corresponding activation function; W i and b i Represents the input gate weight matrix and bias parameters; Based on the forget gate and the input gate, the internal state C at the current moment can be calculated t , and its calculation formula is as follows: Based on the calculation result of the internal state at the current moment, the result obtained under the activation function is combined with the output gate o t Performing Hadamard product operation can obtain the final output h at the current moment t , the calculation formula is as follows: the t =σ(W o ·[h t-1 ,x t ]+b o ) h t =o t ·tanh(C t ) Where W o and b o Represent the output gate weight matrix and bias parameters respectively; Finally, the output results of extracting temporal characteristics and spatial correlation are obtained, and together with the economic indicators for evaluating electricity consumption habits, they are used as the input variables of the fully connected layer. The output result of the fully connected layer is the utilization rate of meteorological sensitive loads in the corresponding period.

6. A system for generating provincial power grid load time sequence scenarios under critical weather conditions for supply guarantee, applied to the method for generating provincial power grid load time sequence scenarios under critical weather conditions for supply guarantee as described in claim 1, characterized in that: include: Analysis module: It is used to propose basic load and meteorological sensitive load by integrating provincial load power and historical meteorological data of load center, and divide meteorological sensitive load into two parts: meteorological sensitive load utilization rate and meteorological sensitive load peak capacity; Grey system model estimation module: used to estimate the base load and meteorological sensitive load peak capacity in the future target year using the grey system model; Deep learning model estimation module: used to combine the CNN convolutional neural network that reflects spatial information and the LSTM model that can extract key information of long time series to generate the meteorological sensitive load utilization rate in the future target year; Generation module: used to multiply the utilization rate of the meteorological sensitive load in the future target year by the peak capacity of the meteorological sensitive load in the future target year to obtain the generated meteorological sensitive load. The generated meteorological sensitive load plus the basic load in the future target year is the generated provincial load power, that is, to generate the provincial power grid load timing scenario under the key weather process for supply guarantee.

Images