Distribution network planning index prediction method based on artificial intelligence

Abstract

The invention relates to the technical field of distribution network planning, and provides a distribution network planning index prediction method based on artificial intelligence, which combines three models to establish a multi-level model, and significantly improves the accuracy and efficiency of distribution network planning index prediction. The main trend and influence factors of distribution network planning are identified by using a multi-layer sensor (MLP), and the nonlinear modeling capability is provided. The time dependence and periodic change of data are captured through a time sequence model ARIMA, and the sensitivity to short-term fluctuation is ensured. A long short-term memory (LSTM) network captures a complex non-linear relationship and long-term dependence, and the prediction precision is further improved. And finally, through a Blending technology, weighted summation is carried out on prediction results of the models, so that maximum output of model advantages is realized, and the accuracy of a fusion result is optimized. Model parameters are dynamically adjusted through a real-time data flow mechanism and an adaptive learning algorithm. Therefore, the prediction is real-time, the method can adapt to continuously changing market and environment conditions, and high prediction precision is kept.

Description

Technical Field

[0001] The present invention relates to the technical field of distribution network planning, and in particular to a distribution network planning index prediction method based on artificial intelligence. Background Art

[0002] Distribution network planning aims to design and optimize the power distribution network to ensure efficient, safe and reliable power supply. At present, with the continuous growth of power demand and the increasing penetration of renewable energy power generation such as wind power generation and photovoltaic power generation in the distribution network, it has brought great difficulties to the planning and management of the distribution network, and it is necessary to study more effective distribution network planning methods.

[0003] Distribution network planning indicator prediction is a process of predicting the future demand and performance of the power distribution network by analyzing historical and real-time data and combining environmental and economic factors. Its core is to evaluate the capacity and efficiency of the existing power grid to ensure that it can meet future load growth and electricity demand. It can help power companies to rationally allocate resources, optimize the layout and scale of power facilities, avoid waste of resources, and identify potential power supply bottlenecks and weaknesses in advance, so as to take measures to improve the stability and reliability of power supply and ensure that the development of the power grid is coordinated with urban planning and economic growth. Machine learning technology has made great progress in recent years, so the current prediction of distribution network planning indicators mainly focuses on machine learning methods. At present, there are the following problems in the prediction of distribution network planning indicators: the accuracy and efficiency of distribution network planning indicator prediction are low, and the data sources are inconsistent: different indicators rely on different data sources, and inconsistent data sources will lead to differences in indicator calculation results, affecting the accuracy of decision-making; lack of dynamic adjustment mechanism: distribution network planning indicators are usually set based on historical data and experience, but with the passage of time and changes in the environment, these indicators will undergo significant changes; insufficient correlation between indicators: the correlation between indicators is not taken into account when setting them, which will lead to unreasonable planning; external factors are not considered: when setting indicators, the impact of external factors such as economic fluctuations and climate change is not considered.

[0004] Therefore, how to overcome the shortcomings of existing technologies and improve the ability of distribution network planning indicators to predict and evaluate existing power grids is an urgent problem to be solved in the field of distribution network planning technology. Summary of the invention

[0005] The purpose of the present invention is to solve the deficiencies of the prior art and to provide a distribution network planning index prediction method based on artificial intelligence.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] A distribution network planning indicator prediction method based on artificial intelligence includes the following steps:

[0008] Step (1), data collection, collecting n types of distribution network planning related element data;

[0009] Step (2), data preprocessing, removing outliers, cleaning data, denoising, filling missing values, and processing by standardization or normalization on the n types of distribution network planning related element data collected in step (1);

[0010] Step (3), model establishment and dynamic adjustment, establish a multi-level comprehensive model, combine the multi-layer perceptron MLP, time series model ARIMA and long short-term memory LSTM network to predict the distribution network planning indicators; combine the predicted values ​​of the three models for weighted summation processing, and realize dynamic adjustment of model parameters through real-time data flow mechanism and adaptive learning algorithm, including the following steps:

[0011] Step (3.1), first level model: use multi-layer perceptron MLP to identify the main trends and influencing factors;

[0012] Step (3.2), second level model: ARIMA analysis is used to capture the time dependence and periodic changes in the data;

[0013] Step (3.3), the third level model: processing complex and nonlinear relationships through the long short-term memory LSTM network model to improve the prediction accuracy;

[0014] Step (3.4), weighted summation processing: Use blending technology to perform fusion through the weighted summation formula:

[0015]

[0016] Among them t is the prediction output of the LSTM network model, Y t is the time series model ARIMA output, is the MLP prediction output; α1, α2, and α3 are weight coefficients between 0 and 1, which are determined by the grid search method to ensure the accuracy of the output;

[0017] Step (3.5), dynamic adjustment mechanism: establish a real-time data flow mechanism, set trigger conditions, and start the adaptive learning algorithm according to significant changes to dynamically adjust the model parameters;

[0018] Step (4), model evaluation and optimization, using cross-validation method to evaluate the model and optimize the model hyperparameters based on the evaluation results;

[0019] Step (5), indicator prediction, uses the optimized model to predict the distribution network planning indicators. The prediction steps include:

[0020] Step (5.1), process the model input data: the historical data of n types of distribution network planning related elements are processed according to step (2)

[0021] Process the data into the format required by the model;

[0022] Step (5.2), perform prediction: load the trained multi-level comprehensive model in step (3), input the model input data to obtain the comprehensive prediction result;

[0023] Step (5.3), processing the prediction results: denormalizing the comprehensive prediction results obtained in step (5.2) to ensure the accuracy of the results, and storing the denormalized prediction results for subsequent analysis;

[0024] Step (6), result analysis and visualization, uses visualization tools to display the forecast results, helping planners understand load change trends and make distribution network planning recommendations.

[0025] Preferably, in the step (1), n=5 in the n types of distribution network planning related element data; the 5 types of distribution network planning related element data are: power load, monthly power consumption, temperature, rainfall, GDP growth rate, population growth rate, transformer data, line data, wind power generation, and solar power generation.

[0026] Preferably, the data preprocessing in step (2) is performed to filter out outliers using the following formula:

[0027]

[0028] Among them, Z is the standard score, μ is the mean, and σ is the standard deviation. When Z exceeds the preset threshold, the corresponding data is judged as an outlier and removed;

[0029] The data preprocessing in step (2) is to screen out outliers, clean and denoise the data, fill in missing values, and normalize the data using the following formula:

[0030]

[0031] X' is the normalized value, X is the original data value, and X min and X max are the minimum and maximum values ​​respectively.

[0032] Preferably, the first level model in step (3.1) identifies the main trends and influencing factors in the data through the following MLP analysis:

[0033]

[0034] The estimation of the regression coefficients is obtained by minimizing the following loss function:

[0035]

[0036] in: is the predicted output load forecast; σ is the nonlinear activation function; β0 is the bias; β1,β2,…,β n are weights; x1,x2,…,x n is the input feature electricity load, monthly electricity consumption, temperature, rainfall, GDP growth rate, population growth rate, transformer data, line data, wind power generation, solar power generation, etc.; m is the number of samples, y i is the actual value, is the predicted value.

[0037] Preferably, the second level model in step (3.2) captures the time dependence and periodic changes in the data through the time series model ARIMA, specifically:

[0038] Y t =Δ d Y t =Y t -Y t-1 -…-Y t-d

[0039] Y t =c+φ1Y t-1 +φ2Y t-2 +…+φ p Y -p +θ1∈ t-1 +θ2∈ t-2 +…+θ q ∈ t-q +∈ t

[0040] Where: Y t is the predicted value at time t; c is a constant term; φ i is the coefficient of the autoregressive part; θ i is the coefficient of the moving average part; ∈ t is white noise; p is the autoregressive order, d is the difference order, and q is the moving average order.

[0041] Preferably, the third-level model in step (3.3) is used to capture the long-term dependencies in the sequence data by establishing a long short-term memory LSTM network model, and the state update of the LSTM network is realized by the following formula:

[0042] (a), Forget gate:

[0043] f t =σ(W f [h t-1 ,x t ]+bf )

[0044] (b), Input gate:

[0045] i t =σ(W i [h t-1 ,x t ]+b i )

[0046] (c), Candidate memory cells:

[0047]

[0048] (d), memory unit update:

[0049]

[0050] (e,) Output gate:

[0051] o t =σ(W o [h t-1 ,x t ]+b o )

[0052] (f), hidden state update:

[0053] h t =o t *tanh(C t );

[0054] Where σ is the Sigmoid activation function; W f and b f are the weight and bias of the forget gate respectively; W i , W C and b i , b C are the weights and biases of the input gate and candidate memory; W o and b o is the weight and bias of the output gate, h t is the hidden state at the current time step.

[0055] Preferably, the dynamic adjustment mechanism in step (3.5) further comprises the following steps:

[0056] a. Deviation detection: After each prediction cycle, the deviation between the predicted result and the actual value is calculated to evaluate the prediction performance of the model. The specific calculation formula is:

[0057]

[0058] in: is the prediction result, ynew (i) is the actual value;

[0059] b. Model adjustment: First, according to the deviation result Deviation, iteratively adjust the parameters of each layer model, and then iteratively adjust the parameters after the model fusion using the Blending technology. The specific steps are as follows:

[0060] (a) MLP weight and bias adjustment: Use the Adam optimization algorithm to iteratively update the network weights. The specific steps are as follows:

[0061] Compute the gradient:

[0062]

[0063] Update the first- and second-order moment estimates:

[0064] m k =β1m k-1 +(1-β1)g

[0065] v k =β2v k-1 +(1-β2)g⊙g

[0066] Correction deviation:

[0067]

[0068] Update parameters:

[0069]

[0070] Among them, m k and v k are estimates of the first and second moments of the gradient, β1 and β2 are exponential decay rates that control the estimates of these two moments, usually set to 0.9 and 0.999, ∈ is a small constant to prevent division by zero, k is the current iteration number, η is the learning rate, and θ k is the set of weights and biases at the kth iteration of the MLP, which is continuously iterated during the training process;

[0071] (b) Adjustment of ARIMA parameters of time series model: determine the difference order d, autocorrelation coefficient ACF, partial autocorrelation coefficient PACF, obtain the best level p and order q, and then obtain the ARIMA model through d, p, q for parameter optimization:

[0072] Traverse all possible parameter combinations (p, d, q), use the information criterion AIC / BIC to select the best parameter combination; fit the ARIMA model and estimate the parameters, use the Liung-Box test to check the autocorrelation of the model residuals, and continuously adjust the parameter combination (p, d, q) to ensure that the model fits well:

[0073]

[0074] Among them, c is a constant term, is the autoregressive coefficient, θ j is the moving average coefficient, ∈ t is white noise;

[0075] (c) Long short-term memory LSTM network weight and bias adjustment: Use stochastic gradient descent (SGD) to update the LSTM weights and bias:

[0076] Calculate the gradient: uniformly sample a sample index i∈{1,…,n} and calculate the gradient of the loss function L(θ)

[0077] Gradient Update:

[0078]

[0079] Among them, l is the current iteration number, η is the learning rate; θ l is the weight and bias set of the LSTM model at the lth iteration, is the gradient of the loss function L with respect to the set of weights and biases θ, indicating the direction of change of the current parameters;

[0080] (d) Blending model parameter adjustment: SGD is used to further iteratively optimize the model weight parameters obtained by blending the blending technology to optimize the weighted sum prediction result:

[0081]

[0082] Among them, a is the weight parameter of the Blending model.

[0083] Preferably, the model evaluation in step (4) adopts a K-fold cross validation method, which specifically includes the following steps:

[0084] (a) Randomly divide the entire data set D into K subsets of the same size D1, D2, ..., D K ;

[0085] (b) Using one of the subsets D K As the validation set, the remaining K-1 subsets are used as training sets;

[0086] (c), train the model M and calculate the validation set D K Performance indicators on k ;

[0087] (d) Repeat the above steps K times and finally calculate the average value e of all K test results avg , according to e avg The size of the established multi-level comprehensive model was adjusted.

[0088] Preferably, the post-processing prediction results in step (5.3) includes performing inverse normalization processing on the prediction results according to the following formula to ensure the accuracy of the results:

[0089] X=X′(X max -X min )+X min

[0090] Among them, X is the original data value, X' is the normalized value, and X min and X max are the minimum and maximum values ​​respectively.

[0091] The present invention discloses a distribution network planning index prediction method based on artificial intelligence, which has the following technical effects.

[0092] The present invention combines three models to establish a multi-level model, which significantly improves the accuracy and efficiency of distribution network planning indicator prediction. The multi-layer perceptron MLP is used to identify the main trends and influencing factors of distribution network planning, and provides nonlinear modeling capabilities. The time series model ARIMA captures the time dependence and periodic changes of data to ensure sensitivity to short-term fluctuations. The long short-term memory network LSTM focuses on capturing complex nonlinear relationships and long-term dependencies, further improving the accuracy of prediction. Finally, through the blending technology, the prediction results of each model are weighted and summed, achieving the maximum output of the model advantages and optimizing the accuracy of the fusion results. Through the real-time data flow mechanism and adaptive learning algorithm, the model parameters can be dynamically adjusted. The prediction is not only real-time, but also adaptable to the changing market and environmental conditions, maintaining high prediction accuracy. The present invention achieves accurate prediction of distribution network planning indicators through the complementary advantages of multi-level models and combined with a dynamic adjustment mechanism. It not only helps power companies to rationally allocate resources and optimize facility layout, but also can identify potential power supply bottlenecks in advance, improve power supply stability and reliability, and ensure that power grid development is coordinated with urban planning and economic growth. BRIEF DESCRIPTION OF THE DRAWINGS

[0093] Figure 1 The present invention is a flow chart of the distribution network planning index prediction method based on artificial intelligence.

[0094] Figure 2 It is a structural schematic diagram of the multi-level comprehensive model of the present invention. DETAILED DESCRIPTION

[0095] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0096] Example

[0097] A distribution network planning indicator prediction method based on artificial intelligence includes the following steps:

[0098] Step (1), data collection, collecting n types of distribution network planning related element data;

[0099] In step (1), n=5 in the n types of distribution network planning related element data; the 5 types of distribution network planning related element data are: power load, monthly power consumption, temperature, rainfall, GDP growth rate, population growth rate, transformer data, line data, wind power generation, and solar power generation.

[0100] Step (2), data preprocessing, in order to improve data quality and data integrity, outliers are removed, data is cleaned, denoised, missing values ​​are filled, and data is processed by standardization or normalization;

[0101] The data preprocessing in step (2) screens out outliers using the following formula:

[0102]

[0103] Among them, Z is the standard score, μ is the mean, and σ is the standard deviation. When Z exceeds the preset threshold, the corresponding data is judged as an outlier and removed;

[0104] In order to enhance the stability and efficiency of model training, the data preprocessing in step (2) removes outliers, cleans and denoises the data, fills in missing values, and normalizes the data using the following formula:

[0105]

[0106] X' is the normalized value, X is the original data value, and X min and X max are the minimum and maximum values ​​respectively.

[0107] Step (3), model establishment and dynamic adjustment, establish a multi-level comprehensive model, combine the multi-layer perceptron MLP, time series model ARIMA and long short-term memory LSTM network to predict the distribution network planning indicators; combine the predicted values ​​of the three models for weighted summation processing, and realize dynamic adjustment of model parameters through real-time data flow mechanism and adaptive learning algorithm, including the following steps:

[0108] Step (3.1), first level model: use multi-layer perceptron MLP to identify the main trends and influencing factors;

[0109] The first level model in step (3.1) identifies the main trends and influencing factors in the data through the following MLP analysis:

[0110]

[0111] The estimation of the regression coefficients is obtained by minimizing the following loss function:

[0112]

[0113] in: is the predicted output load forecast; σ is the nonlinear activation function; β0 is the bias; β1,β2,…,β n are weights; x1,x2,…,x n is the input feature electricity load, monthly electricity consumption, temperature, rainfall, GDP growth rate, population growth rate, transformer data, line data, wind power generation, solar power generation, etc.; m is the number of samples, y i is the actual value, is the predicted value.

[0114] Step (3.2), second level model: ARIMA analysis is used to capture the time dependence and periodic changes in the data;

[0115] The second-level model in step (3.2) captures the time dependence and cyclical changes in the data through the time series model ARIMA, specifically:

[0116] Y t =Δ d Y t =Y t -Y t-1 -…-Y t-d

[0117] Y t =c+φ1Y t-1 +φ2Y t-2 +…+φ p Y -p +θ1∈ t-1 +θ2∈t-2 +…+θ q ∈ t-q +∈ t

[0118] Where: Y t is the predicted value at time t; c is a constant term; φ i is the coefficient of the autoregressive part; θ i is the coefficient of the moving average part; ∈ t is white noise; p is the autoregressive order, d is the difference order, and q is the moving average order.

[0119] Step (3.3), the third level model: processing complex and nonlinear relationships through the long short-term memory LSTM network model to improve the prediction accuracy;

[0120] The third-level model in step (3.3) is used to capture the long-term dependencies in the sequence data by establishing a long short-term memory LSTM network model. The state update of the LSTM network is implemented by the following formula:

[0121] (a), Forget gate:

[0122] f t =σ(W f [h t-1 ,x t ]+b f )

[0123] (b), Input gate:

[0124] i t =σ(W i [h t-1 ,x t ]+b i )

[0125] (c), Candidate memory cells:

[0126]

[0127] (d), memory unit update:

[0128]

[0129] (e,) Output gate:

[0130] o t =σ(W o [h t-1 ,x t ]+b o )

[0131] (f), hidden state update:

[0132] h t =o t *tanh(C t );

[0133] Where σ is the Sigmoid activation function; W f and b f are the weight and bias of the forget gate respectively; W i , W C and b i , b C are the weights and biases of the input gate and candidate memory; W o and b o is the weight and bias of the output gate, h t is the hidden state at the current time step.

[0134] Step (3.4), weighted summation processing: Use blending technology to perform fusion through the weighted summation formula:

[0135]

[0136] Among them t is the prediction output of the LSTM network model, Y t is the time series model ARIMA output, is the MLP prediction output; α1, α2, and α3 are weight coefficients between 0 and 1, which are determined by the grid search method to ensure the accuracy of the output;

[0137] Multilayer perceptron (MLP) is used to identify the main trends and influencing factors of distribution network planning and provide nonlinear modeling capabilities. The time series model ARIMA is used to capture the time dependency and periodic changes of data to ensure sensitivity to short-term fluctuations. The long short-term memory network (LSTM) focuses on capturing complex nonlinear relationships and long-term dependencies to further improve the accuracy of predictions. Finally, the prediction results of each model are weighted and summed through the blending technology to maximize the output of the model's advantages and optimize the accuracy of the fusion results.

[0138] Step (3.5), dynamic adjustment mechanism: establish a real-time data flow mechanism, set trigger conditions, and start the adaptive learning algorithm according to significant changes to dynamically adjust the model parameters;

[0139] The dynamic adjustment mechanism in step (3.5) further includes the following steps:

[0140] a. Deviation detection: After each prediction cycle, the deviation between the predicted result and the actual value is calculated to evaluate the prediction performance of the model. The specific calculation formula is:

[0141]

[0142] in: is the prediction result, y new (i) is the actual value;

[0143] b. Model adjustment: First, according to the deviation result Deviation, iteratively adjust the parameters of each layer model, and then iteratively adjust the parameters after the model fusion using the Blending technology. The specific steps are as follows:

[0144] (a) MLP weight and bias adjustment: Use the Adam optimization algorithm to iteratively update the network weights. The specific steps are as follows:

[0145] Compute the gradient:

[0146]

[0147] Update the first- and second-order moment estimates:

[0148] m k =β1m k-1 +(1-β1)g

[0149] v k =β2v k-1 +(1-β2)g⊙g

[0150] Correction deviation:

[0151]

[0152] Update parameters:

[0153]

[0154] Among them, m k and v k are estimates of the first and second moments of the gradient, β1 and β2 are exponential decay rates that control the estimates of these two moments, usually set to 0.9 and 0.999, ∈ is a small constant to prevent division by zero, k is the current iteration number, η is the learning rate, and θ k is the set of weights and biases at the kth iteration of the MLP, which is continuously iterated during the training process;

[0155] (b) Adjustment of ARIMA parameters of time series model: determine the difference order d, autocorrelation coefficient ACF, partial autocorrelation coefficient PACF, obtain the best level p and order q, and then obtain the ARIMA model through d, p, q for parameter optimization:

[0156] Traverse all possible parameter combinations (p, d, q), use the information criterion AIC / BIC to select the best parameter combination; fit the ARIMA model and estimate the parameters, use the Liung-Box test to check the autocorrelation of the model residuals, and continuously adjust the parameter combination (p, d, q) to ensure that the model fits well:

[0157]

[0158] Among them, c is a constant term, is the autoregressive coefficient, θ j is the moving average coefficient, ∈ t is white noise;

[0159] (c) Long short-term memory LSTM network weight and bias adjustment: Use stochastic gradient descent (SGD) to update the LSTM weights and bias:

[0160] Calculate the gradient: uniformly sample a sample index i∈{1,…,n} and calculate the gradient of the loss function L(θ)

[0161] Gradient Update:

[0162]

[0163] Among them, l is the current iteration number, η is the learning rate; θ l is the weight and bias set of the LSTM model at the lth iteration, is the gradient of the loss function L with respect to the set of weights and biases θ, indicating the direction of change of the current parameters;

[0164] (d) Blending model parameter adjustment: SGD is used to further iteratively optimize the model weight parameters obtained by blending the blending technology to optimize the weighted sum prediction result:

[0165]

[0166] Among them, a is the weight parameter of the Blending model.

[0167] Through the real-time data flow mechanism and adaptive learning algorithm, the model parameters can be adjusted dynamically, making the prediction not only real-time, but also adapting to the ever-changing market and environmental conditions and maintaining high prediction accuracy.

[0168] Step (4), model evaluation and optimization, using cross-validation method to evaluate the model and optimize the model hyperparameters based on the evaluation results;

[0169] The model evaluation in step (4) adopts the K-fold cross-validation method, which specifically includes the following steps:

[0170] (a) Randomly divide the entire data set D into K subsets of equal size D1, D2, …, DK;

[0171] (b) Use one of the subsets DK as the validation set and the remaining K-1 subsets as the training set;

[0172] (c), train the model M and calculate the performance index ek on the validation set DK;

[0173] (d) Repeat the above steps K times, and finally calculate the average value eavg of all K test results, and adjust the established multi-level comprehensive model according to the size of eavg.

[0174] Step (5), indicator prediction, uses the optimized model to predict the distribution network planning indicators. The prediction steps include:

[0175] Step (5.1), process the model input data: the historical data of n types of distribution network planning related elements are processed according to step (2)

[0176] Process the data into the format required by the model;

[0177] Step (5.2), perform prediction: load the trained multi-level comprehensive model in step (3), input the model input data to obtain the comprehensive prediction result;

[0178] Step (5.3), processing the prediction results: denormalizing the comprehensive prediction results obtained in step (5.2) to ensure the accuracy of the results, and storing the denormalized prediction results for subsequent analysis;

[0179] The post-processing prediction results in step (5.3) include inverse normalization of the prediction results using the following formula to ensure the accuracy of the results:

[0180] X=X′(X max -X min )+X min

[0181] Among them, X is the original data value, X' is the normalized value, and X min and X max are the minimum and maximum values ​​respectively.

[0182] Step (6), result analysis and visualization, uses visualization tools to display the forecast results, helping planners understand load change trends and make distribution network planning recommendations.

Claims

1. A distribution network planning index prediction method based on artificial intelligence, characterized in that: The steps include: Step (1), data collection, collecting n types of distribution network planning related element data; Step (2), data preprocessing, removing outliers, cleaning data, denoising, filling missing values, and processing by standardization or normalization on the n types of distribution network planning related element data collected in step (1); Step (3), model establishment and dynamic adjustment, establish a multi-level comprehensive model, combine the multi-layer perceptron MLP, time series model ARIMA and long short-term memory LSTM network to predict the distribution network planning indicators; combine the predicted values ​​of the three models for weighted summation processing, and realize dynamic adjustment of model parameters through real-time data flow mechanism and adaptive learning algorithm, including the following steps: Step (3.1), first level model: use multi-layer perceptron MLP to identify the main trends and influencing factors; Step (3.2), second level model: ARIMA analysis is used to capture the time dependence and periodic changes in the data; Step (3.3), the third level model: processing complex and nonlinear relationships through the long short-term memory LSTM network model to improve the prediction accuracy; Step (3.4), weighted summation processing: Use blending technology to perform fusion through the weighted summation formula: Among them t is the prediction output of the LSTM network model, Y t is the time series model ARIMA output, is the MLP prediction output; α1, α2, and α3 are weight coefficients between 0 and 1, which are determined by the grid search method to ensure the accuracy of the output; Step (3.5), dynamic adjustment mechanism: establish a real-time data flow mechanism, set trigger conditions, and start the adaptive learning algorithm according to significant changes to dynamically adjust the model parameters; Step (4), model evaluation and optimization, using cross-validation method to evaluate the model and optimize the model hyperparameters based on the evaluation results; Step (5), indicator prediction, uses the optimized model to predict the distribution network planning indicators. The prediction steps include: Step (5.1), processing model input data: processing the historical data of n types of distribution network planning related elements into the data format required by the model according to step (2); Step (5.2), perform prediction: load the trained multi-level comprehensive model in step (3), input the model input data to obtain the comprehensive prediction result; Step (5.3), processing the prediction results: denormalizing the comprehensive prediction results obtained in step (5.2) to ensure the accuracy of the results, and storing the denormalized prediction results for subsequent analysis; Step (6), result analysis and visualization, uses visualization tools to display the forecast results, helping planners understand load change trends and make distribution network planning recommendations.

2. The distribution network planning index prediction method based on artificial intelligence according to claim 1, characterized in that: In the step (1), n=5 in the n types of distribution network planning related element data; The five types of distribution network planning-related data elements are: power load, monthly power consumption, temperature, rainfall, GDP growth rate, population growth rate, transformer data, line data, wind power generation, and solar power generation.

3. The distribution network planning index prediction method based on artificial intelligence according to claim 1, characterized in that: The data preprocessing in step (2) is performed to filter out outliers using the following formula: Among them, Z is the standard score, μ is the mean, and σ is the standard deviation. When Z exceeds the preset threshold, the corresponding data is judged as an outlier and removed; The data preprocessing in step (2) is to screen out outliers, clean and denoise the data, fill in missing values, and normalize the data using the following formula: X' is the normalized value, X is the original data value, and X min and X max are the minimum and maximum values ​​respectively.

4. The distribution network planning index prediction method based on artificial intelligence according to claim 1, characterized in that: The first level model in step (3.1) identifies the main trends and influencing factors in the data through the following MLP analysis: The estimation of the regression coefficients is obtained by minimizing the following loss function: in: is the predicted output load forecast; σ is the nonlinear activation function; β0 is the bias; β1,β2,…,β n are weights; x1,x2,…,x n is the input feature electricity load, monthly electricity consumption, temperature, rainfall, GDP growth rate, population growth rate, transformer data, line data, wind power generation, solar power generation; m is the number of samples, y i is the actual value, is the predicted value.

5. The distribution network planning index prediction method based on artificial intelligence according to claim 1, characterized in that: The second-level model in step (3.2) captures the time dependence and periodic changes in the data through the time series model ARIMA, specifically: AND t =Δ d AND t =And t -AND t-1 -…-AND t-d Y t =c+φ1Y t-1 +φ2Y t-2 +…+φ p Yt -p +θ1∈ t-1 +θ2∈ t-2 +…+θ q ∈ t-q +∈ t Where: Y t is the predicted value at time t; c is a constant term; φ i is the coefficient of the autoregressive part; θ i is the coefficient of the moving average part; ∈ t is white noise; p is the autoregressive order, d is the difference order, and q is the moving average order.

6. The distribution network planning index prediction method based on artificial intelligence according to claim 1, characterized in that: The third-level model in step (3.3) is used to capture the long-term dependencies in the sequence data by establishing a long short-term memory LSTM network model. The state update of the LSTM network is realized by the following formula: (a), Forget gate: f t =σ(W f [h t-1 ,x t ]+b f ) (b), Input gate: i t =σ(W i [h t-1 ,x t ]+b i ) (c), Candidate memory cells: (d), memory unit update: (e,) Output gate: the t =σ(W o [h t-1 ,x t ]+b o ) (f), hidden state update: h t =o t *tanh(C t ); Where σ is the Sigmoid activation function; W f and b f are the weight and bias of the forget gate respectively; W i , W C and b i , b C are the weights and biases of the input gate and candidate memory; W o and b o is the weight and bias of the output gate, h t is the hidden state at the current time step.

7. The distribution network planning index prediction method based on artificial intelligence according to claim 1 is characterized in that: The dynamic adjustment mechanism in step (3.5) further comprises the following steps: a. Deviation detection: After each prediction cycle, the deviation between the predicted result and the actual value is calculated to evaluate the prediction performance of the model. The specific calculation formula is: in: is the prediction result, y new (i) is the actual value; b. Model adjustment: First, according to the deviation result Deviation, iteratively adjust the parameters of each layer model, and then iteratively adjust the parameters after the model fusion using the Blending technology. The specific steps are as follows: (a) MLP weight and bias adjustment: Use the Adam optimization algorithm to iteratively update the network weights. The specific steps are as follows: Compute the gradient: Update the first- and second-order moment estimates: m k =β1m k-1 +(1-β1)g v k =β2v k-1 +(1-β2)g⊙g Correction deviation: Update parameters: Among them, m k and v k are estimates of the first and second moments of the gradient, β1 and β2 are exponential decay rates that control the estimates of these two moments, usually set to 0.9 and 0.999, ∈ is a small constant to prevent division by zero, k is the current iteration number, η is the learning rate, and θ k is the set of weights and biases at the kth iteration of the MLP, which is continuously iterated during the training process; (b) Adjustment of ARIMA parameters of time series model: determine the difference order d, autocorrelation coefficient ACF, partial autocorrelation coefficient PACF, obtain the best level p and order q, and then obtain the ARIMA model through d, p, q for parameter optimization: Traverse all possible parameter combinations (p, d, q), use the information criterion AIC / BIC to select the best parameter combination; fit the ARIMA model and estimate the parameters, use the Liung-Box test to check the autocorrelation of the model residuals, and continuously adjust the parameter combination (p, d, q) to ensure that the model fits well: Among them, c is a constant term, is the autoregressive coefficient, θ j is the moving average coefficient, ∈ t is white noise; (c) Long short-term memory LSTM network weight and bias adjustment: Use stochastic gradient descent (SGD) to update the LSTM weights and bias: Calculate the gradient: uniformly sample a sample index i∈{1,…,n} and calculate the gradient of the loss function L(θ) Gradient Update: Among them, l is the current iteration number, η is the learning rate; θ l is the weight and bias set of the LSTM model at the lth iteration, is the gradient of the loss function L with respect to the set of weights and biases θ, indicating the direction of change of the current parameters; (d) Blending model parameter adjustment: SGD is used to further iteratively optimize the model weight parameters obtained by blending the blending technology to optimize the weighted sum prediction result: Among them, a is the weight parameter of the Blending model.

8. The distribution network planning index prediction method based on artificial intelligence according to claim 1, characterized in that: The model evaluation in step (4) adopts the K-fold cross validation method, which specifically includes the following steps: (a) Randomly divide the entire data set D into K subsets of the same size D1, D2, ..., D K ; (b) Using one of the subsets D K As the validation set, the remaining K-1 subsets are used as training sets; (c), train the model M and calculate the validation set D K Performance indicators on k ; (d) Repeat the above steps K times and finally calculate the average value e of all K test results avg , according to e avg The size of the established multi-level comprehensive model was adjusted.

9. The distribution network planning index prediction method based on artificial intelligence according to claim 1, characterized in that: The post-processing prediction results in step (5.3) include performing inverse normalization processing on the prediction results according to the following formula to ensure the accuracy of the results: X=X′(X max -X min )+X min Among them, X is the original data value, X' is the normalized value, and X min and X max are the minimum and maximum values ​​respectively.

Images