Low-voltage distributed photovoltaic high-permeability region short-term load prediction method considering weak measurement

Abstract

The invention provides a low-voltage distributed photovoltaic high-permeability region short-term load prediction method considering weak measurement. The method comprises the following steps: constructing a multi-dimensional time sequence feature comprising a lag feature, a periodic feature and a calendar feature, and carrying out special processing on a special leap year condition; historical load data are collected, whether loads contain distributed photovoltaic output or not is recognized, and data labels are added; based on the multi-dimensional time sequence features and the distributed photovoltaic identification features, constructing a prediction model comprising a bidirectional long-short-term memory neural network; carrying out iterative optimization on the key hyper-parameters of the neural network prediction model by adopting a particle swarm optimization algorithm; and performing power load prediction by using the optimized model. According to the invention, through the multi-dimensional time sequence characteristics and the distributed photovoltaic identification characteristics, the load prediction precision is significantly improved, and the related load prediction problem caused by distributed photovoltaic grid connection is effectively solved.

Description

Technical Field

[0001] This application relates to the field of power system load forecasting technology, and in particular to a method for short-term load forecasting in areas with high penetration of low-voltage distributed photovoltaic power generation, taking into account weak measurement capabilities. Background Technology

[0002] Power load forecasting is a key technology for ensuring the safe and stable operation of the power grid and optimizing dispatch decisions. With the rapid development of distributed energy resources, especially distributed photovoltaic (PV), their output is intermittent, fluctuating, and has anti-peak-shaving characteristics, which leads to significant changes in the shape of the power grid's net load curve, posing a serious challenge to the accuracy of traditional load forecasting methods.

[0003] Existing load forecasting methods mainly include traditional statistical methods (such as regression analysis and time series analysis) and artificial intelligence methods (such as support vector machines (SVM), random forests (RF), and long short-term memory networks (LSTM). Traditional statistical methods are ineffective in handling nonlinear features. Although artificial intelligence methods can handle nonlinear problems, they still have limitations when dealing with complex loads containing distributed photovoltaic (PV) power: 1) They fail to effectively distinguish whether the load data includes PV output and its impact; 2) They lack the ability to fuse and represent multi-dimensional time-series characteristics such as load periodicity, lag, and calendar effects; 3) Model hyperparameters (such as the number of LSTM units and the learning rate) have a significant impact on prediction accuracy, but manual parameter tuning is inefficient and it is difficult to find the optimal solution.

[0004] Therefore, there is an urgent need for a load forecasting method that can effectively integrate multi-dimensional time-series characteristics, accurately identify the impact of distributed photovoltaics, and automatically optimize model parameters in order to improve the load forecasting accuracy under the new power system. Summary of the Invention

[0005] In view of this, this application provides a short-term load forecasting method for low-voltage distributed photovoltaic high-penetration areas that takes into account weak measurement, aiming to solve the problem of decreased forecasting accuracy caused by insufficient utilization of multi-dimensional time series characteristics and the complexity of load patterns caused by distributed photovoltaic grid connection in the prior art.

[0006] This application provides a method for short-term load forecasting in areas with high penetration of low-voltage distributed photovoltaic power generation, considering weak measurement capabilities, including:

[0007] Construct multidimensional time series features, which include lag features, periodic features and calendar features, and perform special processing for leap years;

[0008] Four feature parameters are extracted based on the net load data to identify whether the load data includes distributed photovoltaic output, and corresponding distributed photovoltaic identification tags are added to the historical load data.

[0009] Based on the aforementioned multidimensional temporal features and distributed photovoltaic identification tags, a prediction model containing a bidirectional long short-term memory neural network is constructed.

[0010] An optimization algorithm is used to iteratively optimize the key hyperparameters of the neural network prediction model to obtain the optimal hyperparameter combination.

[0011] The neural network prediction model is trained using the optimal hyperparameter combination to obtain the final load prediction model;

[0012] The final load forecasting model is used to predict the target load and to perform a comparative analysis with related models.

[0013] Furthermore, a multi-dimensional time series feature is constructed, which includes lag features, periodic features, and calendar features, and special processing is performed for leap years, and corresponding data tags are added to the historical power load data;

[0014] Constructing lag features: Load data corresponding to the historical time point with the highest correlation to the load on the predicted target day were selected as lag features. Pearson correlation analysis, Spearman correlation analysis, and maximum cross-correlation analysis were used to analyze the correlation between multiple historical time points and the load on the predicted target day. The historical time point with the highest correlation to the load on the predicted target day was determined to be: L one year ago. 1y One week ago L 1w And L a day ago 1d This forms the hysteresis characteristic matrix L. t =[L 1y L 1w L 1d ];

[0015] Constructing periodic characteristics: Using trigonometric functions to characterize intraday periodic fluctuations (t) d1 =sin(2π*h / 24),t d2 =cos(2π*h / 24)) and monthly cycle characteristics (t) m =sin(2π*m / 12));

[0016] Construct calendar features: Construct date type labels, including weekday labels, weekend / holiday labels, and specific number of days in a year labels; Label matrix D = [D w I w D y ], where D w Indicates the day of the week (1-7), I w Indicates whether it is a weekend (0 or 1), D y Indicates the specific number of days in a year (1-365 / 366);

[0017] To address the issue of missing data from the previous year on February 29th due to leap years (such as 2024), a mapping function is constructed to handle this problem, using historical power load data from March 1st as a substitute.

[0018] Furthermore, four feature parameters are extracted based on the net load data to identify whether the load data includes distributed photovoltaic output, and corresponding distributed photovoltaic identification tags are added to the historical load data;

[0019] Based on the user's historical net load data, calculate and obtain net load curves under various preset weather conditions; based on the user's historical net load data P net (t)=P real (t)-P pv (t) is used to calculate the net load curves under four preset weather types—sunny, partly cloudy, cloudy, and rainy. Specifically, the calculation method involves averaging the net load data for the corresponding date under each weather type.

[0020] Based on the net load curve, four specified characteristic parameters for determining whether distributed photovoltaic power should be installed are extracted within a preset study period. For the net load curves of sunny and cloudy weather, key characteristic parameters are extracted within a preset study period (09:00 to 16:00), including: a) the ratio of total net load power on sunny and cloudy days (C1); b) the standard value of the depression (C2); c) the depression index (C3); d) the load change rate (C4).

[0021] Based on the extracted feature parameters, a recognition model is constructed using the random forest algorithm to determine whether the user has installed distributed photovoltaic power. The extracted feature parameters C1, C2, C3, and C4 are used to form a feature matrix, which is then input into a preset random forest algorithm for training and judgment. The algorithm outputs whether the user's historical power load data includes a label indicating distributed photovoltaic power output.

[0022] Furthermore, based on the multidimensional time-series features and the distributed photovoltaic identification tags, a prediction model incorporating a bidirectional long short-term memory (LSTM) neural network is constructed. The constructed neural network prediction model uses the bidirectional LSTM neural network as its core structure. The bidirectional LSTM neural network simultaneously learns the forward and backward dependencies of the time series through forward and backward LSTM layers, and its hidden layer states are formed by concatenating forward and backward LSTM states.

[0023] Furthermore, an optimization algorithm is used to iteratively optimize the key hyperparameters of the neural network prediction model to obtain the optimal combination of hyperparameters. The model hyperparameters include the number of bidirectional long short-term memory units, learning rate, Dropout rate, batch size, maximum number of training epochs, learning rate reduction period, and learning rate reduction ratio. The key hyperparameters that need to be optimized include the number of bidirectional long short-term memory units, learning rate, Dropout rate, and batch size.

[0024] The particle swarm optimization algorithm parameters are set as follows: 15 particles, 25 iterations, and 4 optimization dimensions, corresponding to the number of units, learning rate, Dropout rate, and batch size of the bidirectional long short-term memory neural network.

[0025] Set the numerical range for hyperparameter optimization iterations: 64-256 units for bidirectional long short-term memory neural networks, 0.0001-0.01 for learning rate, 0.1-0.5 for dropout rate, and 32-128 for batch size;

[0026] Initialize the particle swarm positions and velocities representing the hyperparameter combinations;

[0027] Iterative optimization: For each particle: a) Configure the bidirectional long short-term memory neural network model using the hyperparameters it represents; b) Train the model on the validation set and calculate the prediction error as the fitness value; c) Update the individual optimal solution and the current global optimal solution for the particle; d) Adjust the particle state according to the standard particle swarm optimization velocity update formula and position update formula.

[0028] The combination of hyperparameters corresponding to the final global optimal solution is output as the optimal hyperparameters.

[0029] Furthermore, the neural network prediction model is trained using the optimal hyperparameter combination to obtain the final load prediction model;

[0030] Using the optimal hyperparameter combination obtained above, a bidirectional long short-term memory neural network model is trained on the complete training set to obtain the final load prediction model.

[0031] Furthermore, the target load is predicted using the final load forecasting model, and a comparative analysis of related models is performed.

[0032] The multidimensional time-series features and distributed photovoltaic identification tags corresponding to the day to be predicted are input into the trained final load prediction model, the predicted load value is output, and a comparative analysis of related models is performed.

[0033] This application has the following technical effects:

[0034] 1. A multi-dimensional time-series feature model containing lag features, periodic features, and calendar features was constructed to fully explore the complex time-dependent patterns of the load;

[0035] 2. By extracting the characteristic parameters (C1, C2, C3, C4) of the net load curve and constructing a distributed photovoltaic identification model using the random forest algorithm, the model can effectively identify whether the load data includes distributed photovoltaic output, providing key discrimination information for the model.

[0036] 3. Employing a bidirectional long short-term memory neural network model, it can simultaneously learn the forward and backward dependencies of the load sequence, thus more effectively capturing complex time series patterns;

[0037] 4. The particle swarm optimization algorithm is used to automatically search for the optimal combination of key hyperparameters of the bidirectional long short-term memory neural network, which significantly improves the model performance.

[0038] 5. Through the combination of the above technologies, experimental verification shows that this method has significant improvements in all evaluation indicators compared with other related models, especially in scenarios with distributed photovoltaics, where the prediction accuracy is higher. Attached Figure Description

[0039] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the accompanying drawings used in the embodiments will be briefly described below. These drawings are incorporated in and constitute a part of this specification. They illustrate embodiments conforming to this disclosure and, together with the specification, serve to explain the technical solutions of this disclosure. It should be understood that the following drawings only show some embodiments of this disclosure and should not be considered as limiting the scope. Those skilled in the art can obtain other related drawings based on these drawings without creative effort.

[0040] Figure 1 This is a flowchart illustrating the overall process of the load forecasting method provided in the embodiments of this application.

[0041] Figure 2 Flowchart for constructing multidimensional time series features;

[0042] Figure 3 Flowchart for feature extraction for distributed photovoltaic identification;

[0043] Figure 4 This is a schematic diagram of the bidirectional long short-term memory neural network model structure;

[0044] Figure 5 A flowchart for optimizing the hyperparameters of a bidirectional long short-term memory neural network for the particle swarm optimization algorithm.

[0045] Figure 6 Correlation indicators at different lag times.

[0046] Figure 7 This is a user's net load curve under four different weather conditions.

[0047] Figure 8 This is a diagram of the smallest unit structure of a long short-term memory neural network.

[0048] Figure 9 This is a diagram of a bidirectional long short-term memory neural network structure.

[0049] Figure 10 (a)-(d) are process diagrams of the particle swarm optimization algorithm.

[0050] Figure 11 This is a graph showing all the prediction results.

[0051] Figure 12 This is a graph showing the prediction results for a certain week.

[0052] Figure 13 A graph showing all predicted results considering distributed photovoltaics.

[0053] Figure 14 This is a chart showing the forecast results for a specific week, taking into account distributed photovoltaic power generation.

[0054] Figure 15 The time-segmented prediction results are shown in the figure. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. The components of the embodiments of this disclosure described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this disclosure provided in the accompanying drawings is not intended to limit the scope of the claimed disclosure, but merely represents selected embodiments of this disclosure. All other embodiments obtained by those skilled in the art based on the embodiments of this disclosure without inventive effort are within the scope of protection of this disclosure.

[0056] like Figure 1 As shown in the embodiments of this application, a short-term load forecasting method for areas with high penetration of low-voltage distributed photovoltaic power generation, considering weak measurement, is provided, including:

[0057] S1: Collect historical power load data and perform data preprocessing;

[0058] All historical power load data for the entire year of 2023, the entire year of 2024, and January to April 2025 were collected. The entire year of 2023 and January to November 2024 were used as the training set, December 2024 as the validation set, and January to April 2025 as the test set. The sampling interval was 15 minutes.

[0059] Preprocessing: Missing values ​​were filled using linear interpolation; the original load data were processed according to x' = (xx min ) / (x max -x min Normalize.

[0060] S2: Constructing multidimensional temporal features;

[0061] according to Figure 6 As shown, Pearson correlation, Spearman correlation, and maximum cross-correlation analysis methods are used to study the correlation of various lag times. The Pearson correlation coefficient primarily measures the linear relationship between two variables, calculated by taking the covariance and the standard deviation of the two variables. The Spearman correlation coefficient is a rank-based nonparametric method used to measure the monotonic relationship between two variables; it first transforms the original data of the variables into ranks, and then calculates the rank-based Pearson correlation coefficient. The maximum cross-correlation coefficient is a nonparametric method used to measure the nonlinear relationship between two variables; it divides the data space into grids and calculates the mutual information in each grid. Comprehensive analysis of the correlation indices at various lag times reveals a high correlation between the current load data and the load data from one year ago, one week ago, and one day ago, with corresponding correlation coefficients of 0.79, 0.79, 0.87; 0.87, 0.87, 0.96; and 0.87, 0.87, 1.00, respectively. Therefore, for the prediction model, the correlation coefficients from one year ago (L...) are... 1y One week ago L 1w And one day ago L 1d The load data is used as the basis for predicting the load of the day, and the lag features in the multi-dimensional time series features are constructed to form the lag feature matrix L. t =[L 1y L 1w L 1d ].

[0062] Using trigonometric functions t d1 =sin(2π*h / 24),t d2 =cos(2π*h / 24) represents the intraday periodic fluctuation and t m =sin(2π*m / 12) characterizes the monthly cycle.

[0063] Construct a date type label matrix D = [D w I w D y ], where D w Indicates the day of the week (1-7), I w Indicates whether it is a weekend (0 or 1), D y It indicates the specific number of days in a year (1-365 / 366).

[0064] If the forecast date is February 29, 2024, a special leap year date, then the historical power load data from March 1, 2023, which is a year earlier, will be used as a substitute for the lagging characteristic load data.

[0065] S3: Distributed photovoltaic identification characteristics;

[0066] Based on the user's historical net load data P net (t)=P real (t)-P pv (t) calculates the net load curves under four preset weather types—sunny (WT-S), partly cloudy (WT-L), cloudy (WT-C), and overcast / rainy (WT-R). Specifically, the calculation method involves averaging the net load data for the corresponding date under each weather type. NLC X (t)=[∑P net，X (t)] / T X .like Figure 7 As shown.

[0067] To ensure photovoltaic power output throughout the study period, the study was conducted from 09:00 to 16:00 each day, with the starting point set as S and the ending point as E. The coordinates corresponding to the net load curve are S(t) and E(t). s ,NLC(t s )), E(t e ,NLC(t e )), where t s t e At times starting point S and ending point E, four feature parameters are extracted as follows:

[0068] a. The ratio of net load to total power on sunny and cloudy days (C1).

[0069] The net total power load during sunny and cloudy days is taken as a ratio, and this parameter is denoted as C1:

[0070]

[0071] In the formula: |·| is the absolute value of ·.

[0072] b. Standard value for concave depression (C2).

[0073] The degree of concavity in the net load curve, i.e., the standard value of concavity, can be used as a criterion for judgment; this parameter is denoted as C2. A straight line L is drawn connecting the starting point S and the ending point E, and the slope k of L is calculated. L The equation for L can be obtained as follows:

[0074] y L (t)=k L ×(tt s )+NLCS (t s (2)

[0075]

[0076] In the formula: y L (t) is a function of line L with respect to time t.

[0077] At this point, let A be the set of all points on L at time t, i.e., A = {y} L (t)|t∈[t s , t e ]}.

[0078] Since there are points on the net load curve that lie below the straight line L, let's denote these points as set B, i.e., B = {NLC}. S (t)|NLC S (t) <y L (t), t∈[t s , t e The standard value of the concave depression, C2, is:

[0079]

[0080] In the formula: |·| represents the number of elements in sets A and B.

[0081] c. Concavity index (C3).

[0082] The concave standard value cannot fully reflect whether a user has installed distributed photovoltaic (PV) power, because at certain times, daytime electricity consumption may decrease, causing the curve to fall below the straight line L. Therefore, another characteristic parameter, the concave index denoted as C3, is introduced for the concave curve. Generally, the lowest net load point during the study period occurs around noon on a sunny day. The x-coordinate of the point at 12:00 noon is denoted as t. noon The coordinates of this point on the net load curve are (t noon ,NLC(t noon Using 08:00 time point t respectively A and 18:00 time point t B By connecting this point to form lines L1 and L2, the slopes k of the two lines can be determined. L1 and k L2 :

[0083]

[0084] k L1 and k L2 Adding the absolute values ​​together gives k S :

[0085] k S=|k L1 |+|k L2 | (7)

[0086] Similarly, let t be the abscissa of the point on the net load curve at 12:00 noon under cloudy weather. noon Using 08:00 time point t A and 18:00 time point t B Connect L3 and L4, and find the slope k of the two lines. L3 and k L4 for:

[0087]

[0088] Adding the absolute values ​​gives k C for:

[0089] k C =|k L3 |+|k L4 | (10)

[0090] k S With k C The ratio is denoted as C3:

[0091]

[0092] d. Load change rate (C4).

[0093] As night falls, the output of distributed photovoltaic power will decrease to zero, while the net load curve will rise rapidly due to peak electricity consumption. Let t be the coordinate of the point at 12:00 noon. noon The coordinates of this point on the net load curve are (t noon ,NLC(t noon The coordinate of the point at 21:00 at night is t. night The coordinates of this point on the net load curve are (t night ,NLC(t night Let λ be the slope of the straight line connecting this point and the ending point E.

[0094]

[0095] In the formula: λ S and λ C The slopes are for sunny and cloudy weather, respectively.

[0096] Let λ S and λ C The ratio is the load change rate C4:

[0097]

[0098] Four feature parameters were extracted from the user's net load curve, and the resulting feature parameter matrix can be used as the model input:

[0099] C = [C1, C2, C3, C4] (15)

[0100] The random forest algorithm is used to train and test the input feature parameter matrix, and the model output is used to determine whether the user has installed distributed photovoltaic power.

[0101] S4: Construct a bidirectional long short-term memory neural network prediction model;

[0102] Construct a neural network model with a bidirectional long short-term memory neural network as its core. The input layer dimension is equal to the multi-dimensional temporal feature dimension, with 3-dimensional lag features, 3-dimensional periodic features, 3-dimensional calendar features, and 1-dimensional distributed photovoltaic identification features, for a total of 10 feature dimensions.

[0103] The main calculation formula for Long Short-Term Memory (LSTM) neural networks is:

[0104]

[0105] In the formula, f t Let be the vector of the forget gate at time step t; σ is the sigmoid activation function; W f b f Here are the weight matrix and bias vector for the forget gate; h t-1 h t x represents the output values ​​of the state unit at time step t-1 and time step t; t The input value for time step t; i t W is the vector of the input gate at time step t; i b i c represents the weight matrix and bias vector of the input gate; t-1 c t W represents the cell state vectors at time steps t-1 and t; o b o The weight matrix and bias vector represent the candidate cell states; o t W is the vector of the output gate at time step t; o b o These are the weight matrix and bias vector of the output gate.

[0106] Long Short-Term Memory Neural Network Structure, such as Figure 8 As shown.

[0107] The main calculation formula for a bidirectional long short-term memory neural network is:

[0108]

[0109] In the formula, The hidden layer state of the forward long short-term memory neural network at time step t; For positive long short-term memory neural network units; x t For time step t; The hidden layer state of the forward long short-term memory neural network at time step t-1; The hidden layer state of the reverse long short-term memory neural network at time step t; It is a reverse long short-term memory neural network unit; H represents the hidden layer state of the inverse long short-term memory neural network at time step t+1. t Let t represent the hidden layer state of the bidirectional long short-term memory neural network at time step t.

[0110] Bidirectional long short-term memory neural network structure such as Figure 9 As shown.

[0111] S5: Particle swarm optimization algorithm for optimizing hyperparameters;

[0112] The hyperparameters of a bidirectional long short-term memory (BSL-RM) neural network, including the number of units, learning rate, dropout rate, and batch size, are optimized using the particle swarm optimization (PSO) algorithm. The specific steps for optimization using PSO are as follows:

[0113] 1) Define the parameters of the particle swarm optimization algorithm: 15 particles, 25 iterations, and 4 optimization dimensions, corresponding to the number of units, learning rate, Dropout rate, and batch size of the bidirectional long short-term memory neural network.

[0114] 2) Define the hyperparameter search space, that is, determine the hyperparameters to be optimized and their ranges (number of units in the bidirectional long short-term memory neural network, learning rate, dropout rate, and batch size). The number of units in the bidirectional long short-term memory neural network is 64-256, the learning rate is 0.0001-0.01, the dropout rate is 0.1-0.5, and the batch size is 32-128.

[0115] 3) Train the model based on the given hyperparameters and return the evaluation metrics for the validation set;

[0116] 4) Initialize the position and velocity of the particle swarm, with each particle representing a set of hyperparameters;

[0117] 5) Iteratively update the particle swarm optimization, record the results of each iteration, and update the global optimal solution in real time;

[0118] 6) Train the final model using the optimal hyperparameters.

[0119] The optimal hyperparameter combination was finally obtained: 84 units, a learning rate of 0.00054, a dropout rate of 0.12, and a batch size of 123. The optimization process is as follows: Figure 10 As shown in (a)-(d) (from left to right, from top to bottom respectively) Figure 10 a, b, c, d).

[0120] S6: Train the final model;

[0121] The final model is configured using the optimal combination of hyperparameters: 84 units, learning rate of 0.00054, dropout rate of 0.12, batch size of 123, maximum training epochs of 50, learning rate reduction period of 15 epochs, and learning rate reduction ratio of 70%.

[0122] S7: Final load forecast.

[0123] The prediction results of the bidirectional long short-term memory neural network prediction model optimized by the particle swarm optimization algorithm are as follows: Figure 11 and Figure 12 As shown.

[0124] By conducting comparative experiments on relevant models, the evaluation indicators for all models are shown in Table 1.

[0125] Table 1 Evaluation metrics for all models

[0126]

[0127] By comparing the evaluation metrics of each model group, it can be found that the predictive performance of the model improves slightly with the increase of training epochs; when the number of training epochs is the same, the two-layer model shows a slight improvement over the single-layer model. Furthermore, because BiLSTM is better at capturing features of complex temporal sequences, its performance is generally superior to LSTM. After optimizing the hyperparameters of the algorithm using PSO, the MAPE of the PSO-LSTM model is 3.2729%, and R... 2 The MAPE of the PSO-BiLSTM model is 0.92165, representing improvements of 9.42% and 3.12% compared to the single-layer LSTM's 3.6132% and 0.89375, respectively; and improvements of 7.64% and 1.78% compared to the two-layer LSTM's 3.5437% and 0.90553, respectively. The MAPE of the PSO-BiLSTM model is 3.2322%, and R0 is... 2 The PSO-optimized algorithm achieves a performance score of 0.92877, representing improvements of 8.93% and 2.84% compared to the single-layer BiLSTM's 3.549% and 0.90311, respectively; and improvements of 7.74% and 2.42% compared to the two-layer BiLSTM's 3.5035% and 0.90684, respectively. This demonstrates that the PSO-optimized algorithm models exhibit superior prediction performance.

[0128] After adding distributed photovoltaic identification features, the prediction results of the bidirectional long short-term memory neural network prediction model optimized by the particle swarm optimization algorithm are as follows: Figures 13-15 As shown in Table 2, the evaluation indicators for the results are as follows.

[0129] Table S7.2 Model Evaluation Indicators

[0130]

[0131] According to the evaluation metrics, the MAPE and R of PSO-LSTM 2 The MAPE and R of PSO-BiLSTM were 5.1321% and 0.9558, respectively. 2 The improvements in accuracy were 4.3905% and 0.9621, respectively, representing a 14.45% and 6.59% increase over PSO-LSTM. The model's prediction accuracy was also improved to some extent due to the inclusion of photovoltaic-related features in the input.

[0132] This method effectively improves the accuracy of power load prediction in scenarios involving distributed photovoltaics by integrating distributed photovoltaic identification tags and multi-dimensional time-series features (lag, period, calendar) and using particle swarm optimization algorithm to optimize the hyperparameters of bidirectional long short-term memory neural networks, thus providing more reliable technical support for the safe and stable operation of new power systems.

[0133] Finally, it should be noted that the above-described embodiments are merely specific implementations of this disclosure, used to illustrate the technical solutions of this disclosure, and not to limit it. The protection scope of this disclosure is not limited thereto. Although this disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in this disclosure; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this disclosure, and should all be covered within the protection scope of this disclosure. Therefore, the protection scope of this disclosure should be determined by the protection scope of the claims.

Claims

1. A method for short-term load forecasting in areas with high penetration of low-voltage distributed photovoltaic power generation, considering weak measurement capabilities, characterized in that... include: Construct multidimensional time series features, which include lag features, periodic features and calendar features, and perform special processing for leap years; Four feature parameters are extracted based on the net load data to identify whether the historical power load data includes distributed photovoltaic output, and corresponding distributed photovoltaic identification tags are added to the historical load data. Based on the aforementioned multidimensional time-series features and distributed photovoltaic identification tags, a neural network prediction model incorporating bidirectional long short-term memory is constructed. An optimization algorithm is used to iteratively optimize the key hyperparameters of the neural network prediction model to obtain the optimal hyperparameter combination. The neural network prediction model is trained using the optimal hyperparameter combination to obtain the final load prediction model; The final load forecasting model is used to predict the target load and to perform a comparative analysis with related models.

2. The method for short-term load forecasting in areas with high penetration of low-voltage distributed photovoltaic power generation considering weak measurement, as described in claim 1, is characterized in that... Constructing multidimensional time-series features specifically includes: adding corresponding data tags to the historical power load data, including: Constructing lag features: Select load data corresponding to the historical time point with the highest correlation to the load of the predicted target day as lag features. After comparing multiple time points, select the three historical time points with the strongest comprehensive correlation. The historical time points include one year ago, one week ago, and one day ago. Constructing cyclical features: Trigonometric functions are used to represent intraday cyclical fluctuations and monthly cyclical features; Build calendar features: Build date type labels, including weekday labels, weekend / weekend labels, and specific number of days in a year labels; To handle leap years, a mapping function is constructed to address the issue of missing data on special dates.

3. The method for short-term load forecasting in areas with high penetration of low-voltage distributed photovoltaic power generation considering weak measurement, as described in claim 1, is characterized in that... Identifying whether the historical power load data includes distributed photovoltaic output includes: Based on the user's historical net load data, calculate and obtain the net load curves under various preset weather types; Based on the net load curve, four specified characteristic parameters for determining whether distributed photovoltaic power should be installed are extracted within a preset study period. Based on the extracted feature parameters, a recognition model is constructed using the random forest algorithm to determine whether the user has installed distributed photovoltaic power.

4. The method for short-term load forecasting in areas with high penetration of low-voltage distributed photovoltaic power generation considering weak measurement, as described in claim 3, is characterized in that... It also includes: using net load curves to obtain characteristic parameters, and then constructing a distributed photovoltaic identification model, including: The preset weather types include four types: sunny, partly cloudy, cloudy, and rainy. The planned study period is from 09:00 to 16:00; The characteristic parameters include the ratio of net load to total power on sunny and cloudy days (C1), the concave standard value (C2), the concave index (C3), and the load change rate (C4), totaling four types.

5. A method for short-term load forecasting in areas with high penetration of low-voltage distributed photovoltaic power generation, considering weak measurement, as described in claim 1, characterized in that... The neural network prediction model is a bidirectional long short-term memory neural network model, including: Model hyperparameters include the number of bidirectional long short-term memory units, learning rate, dropout rate, batch size, maximum number of training epochs, learning rate reduction period, and learning rate reduction percentage. Key hyperparameters that need optimization include the number of bidirectional long short-term memory units, learning rate, dropout rate, and batch size.

6. A method for short-term load forecasting in areas with high penetration of low-voltage distributed photovoltaic power generation, considering weak measurement, as described in claim 1, characterized in that... The step of iteratively optimizing the key hyperparameters of the neural network prediction model using an optimization algorithm includes: Set the parameters of the particle swarm optimization algorithm, including the number of particles, the number of iterations, and the optimization dimension, which is the number of key hyperparameters that need to be optimized; Define the numerical range for each key hyperparameter in the optimization iteration; Initialize the position and velocity of the particle swarm, with each particle representing a set of hyperparameter combinations; Iterative particle swarm update: For each particle, the neural network prediction model is trained using the hyperparameter combination it represents, the prediction error on the validation set is calculated as the fitness value, the individual optimal solution and the global optimal solution are updated in real time, and the particle velocity and position are adjusted according to the update formula. The final globally optimal solution is taken as the optimal combination of hyperparameters.

7. A method for short-term load forecasting in areas with high penetration of low-voltage distributed photovoltaic power generation, considering weak measurement, as described in any one of claims 1-6, characterized in that, After collecting historical power load data, the data is preprocessed, including handling missing values ​​and normalization.

8. A method for short-term load forecasting in areas with high penetration of low-voltage distributed photovoltaic power generation, considering weak measurement, as described in claim 1, characterized in that... include: Comparative analysis of relevant models The relevant models include: a single-layer long short-term memory neural network model, a single-layer bidirectional long short-term memory neural network model, a two-layer long short-term memory neural network model, and a two-layer bidirectional long short-term memory neural network model; The evaluation metrics for comparative analysis include: mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R²). 2 .

Images