Short-term power load probability prediction method based on improved neural network

Abstract

The invention discloses a short-term power load probability prediction method based on an improved neural network, and belongs to the technical field of power system load prediction. The method comprises the following steps: acquiring and preprocessing load and related data; the method comprises the following steps: constructing a DCS-CNN-BiLSTM-Attention-QR probability prediction model fusing a convolutional neural network, a bidirectional long short-term memory network, an attention mechanism, quantile regression and kernel density estimation; automatically optimizing key hyper-parameters of the model by utilizing a difference creation search algorithm; carrying out model training by adopting a quantile loss function to obtain load prediction values under different quantiles; and finally, generating a probability density curve of the prediction interval through kernel density estimation. According to the method, deterministic point prediction is expanded into probabilistic interval prediction, the uncertainty of load prediction can be quantified, the model precision and generalization ability are improved through an intelligent optimization algorithm, and richer and more reliable decision information is provided for power system scheduling and risk management.

Description

Technical Field

[0001] This invention relates to the field of power load forecasting technology, and specifically to a short-term power load probabilistic forecasting method based on deep neural networks. Background Technology

[0002] In today's society, electricity is the core energy source supporting daily life and the vigorous development of industrial production, and its importance is self-evident. Accurate and reasonable forecasting of electricity load is a necessary prerequisite for maintaining the stable operation of the power system and ensuring the reliability of power supply. Although there are various methods for forecasting electricity load, achieving accurate forecasting still faces many challenges and urgently requires further research and innovation.

[0003] For power load forecasting, Wang Yuan et al. mostly used the elasticity matrix method to construct user demand response assessment models (Wang Yuan, Zhou Ming. Evaluation method of residential users' response potential to time-of-use pricing [J]. Electric Power Construction, 2017, 38(11): 48-54.). Zhang Yanchao introduced elasticity factors for modeling. In order to determine the elasticity coefficient, various environmental factors were considered, and relevant data from questionnaire surveys and regional statistics were used to calculate the elasticity coefficient. (Zhang Yanchao. Research on electricity demand response based on residential heterogeneity [D]. Zhenjiang: Jiangsu University, 2020.) However, this method based on supply and demand equilibrium theory has limitations: on the one hand, it ignores the driving effect of environmental factors such as temperature and humidity on user behavior; on the other hand, the fixed elasticity coefficient cannot dynamically adapt to market fluctuations, resulting in limited prediction accuracy and generalization. In order to overcome these drawbacks, neural network methods for mining user data features have emerged. Ruan Jiaqi et al. used Long Short-Term Memory (LSTM) networks to predict user demand response behavior. (Ruan Jiaqi, Liu Wenxuan, Zhao Junhua, et al. Data-driven retail electricity pricing strategy for demand response [J]. Automation of Electric Power Systems, 2023, 47(7): 133-141.) Neural networks have strong fitting and generalization capabilities, which compensate for the shortcomings of inaccurate elasticity matrix prediction and poor generalization ability. Therefore, they are currently widely used in user demand response modeling. However, current research ignores the correlation between electricity demand and electricity prices before and after. Under the implementation of dynamic electricity pricing mechanism, in order to adjust their electricity consumption behavior more effectively, users must consider the current and previous electricity price differences at the time of electricity consumption. Therefore, there is a potential correlation between a user's electricity demand at a certain moment and the electricity price at the previous and previous moments.

[0004] Given the limitations of the aforementioned forecasting methods in accurately measuring the high degree of uncertainty in electricity load, an increasing number of studies are turning to interval forecasting. Interval forecasting not only provides point estimates of future load but also a forecast interval encompassing the range of possible future loads at a certain confidence level. Quantile regression, as an effective interval forecasting tool, combined with machine learning models, can fully leverage the advantages of both. This combination not only improves forecast accuracy but also provides richer information for power system operation, helping decision-makers better address load uncertainty.

[0005] However, when building probabilistic prediction models, the model parameters such as the learning rate are determined through manual experimental experience, and the selection of these hyperparameters has a significant impact on the final model's performance. Finding suitable hyperparameter values ​​is often a complex and time-consuming process. Different datasets and neural network structures may require different hyperparameter settings, making hyperparameter tuning even more difficult and challenging. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and to disclose a short-term power load probability prediction method based on an improved neural network.

[0007] This invention is achieved through the following technical solution.

[0008] The present invention discloses a short-term power load probability prediction method based on an improved neural network, comprising the following steps:

[0009] Step 1: Collect power load data and determine the input data based on the correlation between power load data and external influencing factors;

[0010] Step 2: Preprocess the input data and divide it into training set data and test set data;

[0011] Step 3: Construct a short-term power load probability prediction model based on a DCS-CNN-BiLSTM-Attention-QR (Differential Creative Search-Convolutional Bidirectional Long Short-Term Memory Network-Attention Mechanism-Quantile Regression) neural network. This model no longer uses traditional mean squared error as a loss function, but instead uses quantile loss instead of the loss function in the CNN-BiLSTM-Attention network. This makes the output of the neural network no longer a single predicted value (point prediction), but can output prediction results at different quantiles through the quantile regression layer, thus constructing a prediction interval. The short-term power load probability prediction model of the DCS-CNN-BiLSTM-Attention-QR neural network consists of an input layer, a convolutional layer, a bidirectional long short-term memory neural network layer, a fully connected layer, a quantile regression layer, a kernel density estimation layer, and an output layer.

[0012] Step 4: Initialize the prediction model parameters and set the initial parameters for the Differentiated Creative Search (DCS) optimization algorithm. Addressing the issue that adjusting neural network hyperparameters often relies on manual experience and is difficult to optimize, the DCS algorithm automatically seeks optimizations. It is specifically designed for optimizing hidden nodes, regularization coefficients, and learning rates in Bidirectional Long Short-Term Memory (BiLSTM) networks. This avoids the blindness of manual parameter tuning and improves the model's fitness and prediction accuracy.

[0013] Step 5: Use the training set data from Step 2 to train the model. Use the root mean square error (RMSE) of the point prediction index as the fitness. Combine the DCS algorithm to optimize the number of hidden nodes, regularization coefficient and learning rate of the BiLSTM hidden layer of the prediction model. Use quantile loss as the loss function to obtain the short-term load probability density prediction model of the trained neural network.

[0014] Step Six: Select The quantiles in (0,1). The predicted values ​​are arranged in ascending order of quantiles to form a power load forecast sequence. This sequence not only contains the predicted value for each quantile, but also implies information about the uncertainty distribution of the power load;

[0015] Step 7: Using the predicted values ​​at different quantiles as input, apply a Gaussian kernel to the predicted sequence. We use a weighted average of local regions near each data point to estimate the prediction interval and probability density curve. By fitting the predicted probability density curve, we achieve the transformation from quantile intervals to continuous probability density functions, which can describe the distribution probability of the load at different times in more detail.

[0016] Preferably, the input data in step one includes load data, electricity price, and time type for the same time period of the previous seven days.

[0017] Preferably, the training set accounts for 80% and the test set accounts for 20% in step two. Input data preprocessing employs the mean square method to handle outliers, and then uses formula (1) to normalize the data, eliminating data differences and controlling the value range to between 0 and 1, where x... max x represents the maximum value in the data. min It is the minimum value in the data.

[0018] (1)

[0019] Preferably, the convolutional layer in step three is responsible for extracting features from the input data, the bidirectional long short-term neural network layer captures bidirectional dependency information in the sequence through network layers in both forward and backward directions, the quantile regression layer predicts the power load at different quantiles, and the kernel density estimation layer realizes the conversion from quantile prediction to probability density function.

[0020] Preferably, in step four, the initial value of the number of hidden nodes in the BiLSTM prediction model is 32, with an optimization range of [0, 100]; the initial value of the regularization coefficient is 1E-4, with an optimization range of [1E-6, 1E-2]; and the initial value of the learning rate is 0.001, with an optimization range of [1E-5, 1E-2]. The DCS algorithm has a population size of 30, a maximum number of iterations of 20, and a ratio of high, medium, and low performers of 1:2:3.

[0021] The DCS algorithm flow is shown below:

[0022] Step 1: Set the initial parameters required by the algorithm, including population size, number of optimization parameters, maximum number of iterations, etc. ngs The value of .

[0023] Step 2: Randomly initialize the population, evaluate the fitness of each individual in the initial population according to the objective function, set the best performing individual as the team leader, and classify the other individuals into three groups (high performers, average performers, and poor performers) according to the set proportions.

[0024] Step 3: By calculating the knowledge acquisition rate and randomized differentiated knowledge, each individual acquires knowledge in a differentiated manner.

[0025] Step 4: Conduct divergent thinking strategies, convergent thinking strategies, and team diversification strategies for high-performing, average-performing, and low-performing team members respectively.

[0026] Step 5: Perform a retrospective evaluation of the iterated population and update the group category and corresponding fitness of each team member.

[0027] Step 6: Determine if the maximum number of iterations has been reached. If the maximum number of iterations has not been reached, return to Step 3; otherwise, the algorithm ends.

[0028] Preferably, the short-term load probability density prediction model in step five is trained as follows: using the RMSE of the point prediction index as the fitness, the number of hidden nodes in the BiLSTM layer, the regularization coefficient, and the learning rate of the model are optimized. The quantile loss is used as the loss function to optimize the parameters of the constructed neural network, thus obtaining the short-term load probability density prediction model.

[0029] Preferably, in step six, the number of quantiles in the quantile regression layer is set to 9, the quantile values ​​range from (0,1), and the interval is 0.1, forming a power load forecast sequence. .

[0030] Preferably, the kernel density estimation method in step seven uses a Gaussian kernel as the kernel function and an adaptive bandwidth. The probability density data of the prediction interval is estimated by performing weighted regression on the predicted values ​​of different quantiles.

[0031] The beneficial effects of this invention include: replacing the loss function of the CNN-BiLSTM-Attention neural network with quantile loss, extending the deterministic prediction method to a probabilistic prediction method, and obtaining the probability density curves of each point in the prediction interval at different time points through kernel density estimation; and optimizing the number of BiLSTM hidden layer nodes, regularization coefficients, and learning rate of the model through the DCS algorithm. Attached Figure Description

[0032] Figure 1 This is a flowchart of the DCS algorithm.

[0033] Figure 2 It is a CNN-BiLSTM-Attention neural network structure.

[0034] Figure 3 This is a flowchart of optimizing the parameters of the CNN-BiLSTM-Attention model based on DCS.

[0035] Figure 4 This is a diagram of the DCS-CNN-BiLSTM-Attention-QR model structure.

[0036] Figure 5 This is the DCS fitness curve.

[0037] Figure 6 The pointwise MAPE error for each model's electricity demand forecast for November 30, 2024.

[0038] Figure 7 The image shows the quantile prediction results of the DCS-CNN-BiLSTM-AT-QR model.

[0039] Figure 8 This is a probability density curve for each sampling point. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be described in more detail below with reference to the accompanying drawings.

[0041] This embodiment describes a short-term power load probabilistic prediction method based on an improved neural network. It optimizes network parameters using quantile loss, combines deep neural networks with quantile regression, extends point prediction to interval prediction, and overcomes the limitation of quantile regression in solving complex nonlinear problems. Kernel density estimation is used to fit the probability density distribution of the prediction results, thereby improving the accuracy of power load prediction.

[0042] Its prediction model consists of seven interconnected parts, combining deterministic deep learning feature extraction with probabilistic quantile regression and utilizing metaheuristic algorithms to address the hyperparameter black box problem. The composition and function of each part of the model are as follows:

[0043] (1) Input and preprocessing layer: In order to capture the periodicity and external correlation of the load, the model selects the load data, electricity price and time type at the same time in the previous seven days as input. The mean square value method is used to process the abnormal data and the data is normalized to eliminate the difference in dimensions.

[0044] (2) Feature extraction layer: Utilizing the unique convolutional computation structure of the convolutional neural network, feature extraction and dimensionality reduction are performed on the input time series data, thereby mining high-dimensional local features from the original multidimensional input data.

[0045] (3) BiLSTM layer: It contains network layers in both forward and backward directions. The forward LSTM processes in chronological order, and the backward LSTM processes in reverse order, thereby capturing bidirectional dependency information in the power load sequence and solving the long-distance dependency problem.

[0046] (4) Attention mechanism layer: Based on the output of the BiLSTM layer, different weights are assigned to the feature values ​​at different time steps, so that the model automatically gathers the key historical information that has the greatest impact on the current prediction time, in order to suppress irrelevant noise.

[0047] (5) Quantile regression layer: Instead of constructing loss functions such as the traditional mean square error, the quantile loss function is adopted, with 9 quantiles set. The output is a sequence containing the predicted values ​​under different confidence levels, reflecting the uncertainty distribution of the load.

[0048] (6) Kernel density estimation layer: Using Gaussian kernel function and adaptive bandwidth, weighted regression is performed on the predicted values ​​of different quantiles to generate probability density curves.

[0049] (7) DCS parameter optimization layer: For the three key hyperparameters of BiLSTM hidden layer, such as number of nodes, regularization coefficient, and learning rate, DCS is used to iteratively optimize each hyperparameter through differentiated knowledge acquisition and dual strategies, thereby avoiding the blindness of manual parameter tuning.

[0050] The short-term power load probabilistic prediction model based on the improved neural network uses the following method:

[0051] (1) Convolutional Neural Network: With its unique convolutional computation structure and deep feedforward characteristics, it has significant effects in feature extraction and dimensionality reduction. Its basic components include convolutional layers, pooling layers and fully connected layers.

[0052] (2) Bidirectional Long Short-Term Memory Neural Network: The bidirectional long short-term memory network is based on the LSTM structure. It captures bidirectional dependency information in the sequence through network layers in both forward and backward directions, further expanding its capability boundary. The forward LSTM processes the data from the beginning of the sequence in chronological order, while the backward LSTM processes the data in reverse from the end of the sequence.

[0053] (3) Attention mechanism: In time series models, the attention mechanism focuses on key information by assigning different weights to different parts of the input sequence. Specifically, the model dynamically adjusts the attention to each time point based on historical data and the current input, thereby capturing the core information in the time series more accurately.

[0054] (4) Differentiated Creation Search Algorithm: The DCS algorithm uses differentiated knowledge acquisition of individuals in the population and two types of innovative thinking strategies to solve optimization problems, and has the characteristics of fast convergence speed and efficient search for optimal solutions. The DCS algorithm enhances decision-making efficiency by adopting a novel dual-strategy method, which mainly includes a differentiated knowledge acquisition process, iterative loops of divergent and convergent thinking, and retrospective evaluation.

[0055] (5) Quantile regression method: The traditional QR method, as one of the linear regression methods, can obtain the load value at different quantiles in the future by performing regression analysis on the historical load and related influencing factor datasets. Compared with point prediction, it can provide more comprehensive prediction information.

[0056] (6) Kernel density estimation method: The result of quantile regression convolutional neural network is used as the input of kernel density estimation method to fit the probability density distribution of the prediction result.

[0057] Step 1: Collect power load data and determine the input data based on the correlation between power load data and external influencing factors; select data from a power user in Jiangxi participating in the spot market settlement trial operation, and select the load data, electricity price, and time type of the same time in the previous seven days as input features based on the influencing factor analysis.

[0058] To verify the rationality of the input feature selection, particularly the impact of electricity price on load, this embodiment introduces Copula theory for correlation analysis. As shown in Table 1, the correlation coefficient between electricity price and electricity demand under the t-Copula function is 0.74, which is much higher than the Pearson linear correlation coefficient (0.43). This indicates a very strong nonlinear correlation between the two, validating the necessity of using electricity price as the input feature and employing a deep neural network with nonlinear mapping capabilities for modeling.

[0059] Table 1

[0060]

[0061] Step 2: Preprocess the data, handle outliers, normalize the data to [0,1], and then split the data into training and test sets;

[0062] Input data preprocessing employs the mean square method to handle outliers, and then uses formula (1) to normalize the data, eliminating data discrepancies and controlling the value range to be between 0 and 1, where x max x represents the maximum value in the data. min It is the minimum value in the data.

[0063] (1)

[0064] The first 80% of the dataset was selected as training data, and the remaining 20% ​​was used as test data. The quantile regression layer was set to have 9 quantiles, with quantile values ​​ranging from (0,1) and intervals of 0.1.

[0065] Step 3: Construct a short-term power load probability prediction model based on DCS-CNN-BiLSTM-Attention-QR neural network. This model consists of an input layer, convolutional layers, a bidirectional long short-term neural network layer, a fully connected layer, a quantile regression layer, a kernel density estimation layer, and an output layer. Its model structure is as follows: Figure 4 As shown, the CNN-BiLSTM-Attention network structure is as follows: Figure 2 As shown.

[0066] Step 4: Initialize the prediction model parameters and set the initial parameters required for the DCS optimization algorithm. In Step 4, the initial value for the number of BiLSTM hidden layer nodes is 32, with an optimization range of [0, 100]; the initial value for the regularization coefficient is 1E-4, with an optimization range of [1E-6, 1E-2]; and the initial value for the learning rate is 0.001, with an optimization range of [1E-5, 1E-2]. The DCS algorithm population size is 30, the maximum number of iterations is 20, and the ratio of high, medium, and low performers is 1:2:3. The DCS optimization algorithm flow is as follows: Figure 1 As shown, the process of optimizing the parameters of the CNN-BiLSTM-Attention model based on DCS is as follows: Figure 3 As shown.

[0067] Step 5: Use the training set data from Step 2 to train the model. Use the RMSE of the point prediction index as the fitness to optimize the number of hidden nodes in the BiLSTM hidden layer, the regularization coefficient and the learning rate of the model. Use the quantile loss as the loss function. The quantile loss is shown in formulas (2) and (3) to obtain the trained short-term load probability density prediction model based on DCS-CNN-BiLSTM-Attention-QR neural network.

[0068] (2)

[0069] (3)

[0070] Where τ represents the quantile, ; Indicates an indicator function.

[0071] Step Six: Input the test data into the trained model to obtain predicted values ​​at different quantiles, forming a power load prediction sequence. It contains the predicted value for each quantile.

[0072] Step 7: The kernel density estimation method uses a Gaussian kernel as the kernel function and an adaptive bandwidth. It predicts the probability density curve by performing weighted regression on the predicted values ​​of different quantiles.

[0073] To verify the effectiveness of the method proposed in this invention, data from a power user in Jiangxi Province participating in a trial run of spot market settlement were selected for the experiment. The dataset contains load data from July 1 to November 30, 2024, with a time interval of 1 hour.

[0074] 1. DCS Algorithm Optimization Process and Parameter Determination

[0075] During model training, the DCS algorithm demonstrated efficient optimization capabilities. Figure 5 The fitness evolution curve of the DCS algorithm on the training set is shown. The fitness value (based on RMSE) of the DCS algorithm decreases with the number of iterations and tends to stabilize after the 13th iteration, indicating that the DCS algorithm has fast convergence speed and high search accuracy in neural network hyperparameter optimization. The optimal parameters of the model determined through this optimization process are shown in Table 2.

[0076] Table 2

[0077]

[0078] 2. Analysis of Point Prediction Results

[0079] The test set data was input into the trained model. The mean absolute percentage error (MAPE) and RMSE of each comparative model on the test set are shown in Table 3.

[0080] Table 3

[0081]

[0082] As shown in Table 3, the proposed DCS-CNN-BiLSTM-Attention model (attention mechanism denoted as AT in the table) outperforms other models in both MAPE and RMSE metrics. Specifically, compared to the CNN-BiLSTM-Attention model, GA-CNN-BiLSTM-Attention model, and PSO-CNN-BiLSTM-Attention model, the MAPE metric in the test set decreased by 1.21%, 0.55%, and 0.41%, respectively; and the RMSE metric in the test set decreased by 2.08, 1.03, and 0.58, respectively. These results demonstrate that the optimal parameters found by the DCS algorithm on the CNN-BiLSTM-Attention model yield better results, and the predictions are closer to the true values.

[0083] To observe the model's predictive performance at different times in more detail, the MAPE error distribution for each sampling point on November 30, 2024, was plotted, as shown below. Figure 6 As shown. Combined with Figure 6 Analysis shows that the relative error of the DCS-CNN-BiLSTM-Attention model is mostly controlled within 5%. Especially during the peak period from sampling point 10 to sampling point 20, the error fluctuation of the comparison model (such as other models such as GA-CNN-BiLSTM-AT) increased significantly, while the model of this invention still maintained a low and stable error level, further verifying the robustness of the model under scenarios with drastic load fluctuations.

[0084] 3. Analysis of Probability Prediction Results

[0085] Based on point prediction, the interval prediction performance of the model is further evaluated. The prediction interval coverage probability (PICP) and the normalized average width of prediction intervals (PINAW) of each model are shown in Table 4.

[0086] Table 4

[0087]

[0088] As shown in Table 4, the PICP of the short-term power load probability model based on the improved neural network in this invention reaches 89.21%, while the PINAW is only 15.21%, both of which are superior to other probability prediction models. Compared with CNN-BiLSTM-AT-QR, GA-CNN-BiLSTM-AT-QR, and PSO-CNN-BiLSTM-AT-QR models, the PICP index in the test set is improved by 4.58%, 2.97%, and 0.58%, respectively; while the PINAW index in the test set is reduced by 3.42%, 1.03%, and 0.42%, respectively. These results indicate that this invention, while maintaining high probability prediction accuracy, has a narrower prediction interval and a more outstanding probability prediction effect.

[0089] Let's analyze the probability prediction results using November 30, 2024 as an example:

[0090] Interval Coverage: The vast majority of actual load values ​​(true values) fall within the forecast interval formed by the 9 quantile curves, such as... Figure 7 As shown. By Figure 7 As can be seen, the actual values ​​closely match the predicted curve, indicating that the prediction interval generated by the model has good coverage.

[0091] Dynamic characteristics: During peak load periods (such as sampling points 10-20), the quantile curves show a divergent trend, and the interval width automatically adjusts with load fluctuations, accurately reflecting the increased uncertainty during peak periods.

[0092] Probability density: The probability density curve generated by kernel density estimation is as follows Figure 8 As shown in the figure, by selecting the curves at sampling points 10, 14, 18, and 22 for observation, it can be seen that the peak positions are highly consistent with the actual load values, which can clearly describe the probability distribution of load occurrence at that moment and effectively capture the true fluctuation characteristics of electricity demand.

[0093] In summary, this invention constructs an efficient probabilistic prediction framework based on an improved neural network-based short-term power load probabilistic prediction method. The model first utilizes CNN and BiLSTM components to deeply mine multidimensional features and bidirectional temporal correlations in power load data, and then uses an attention mechanism to accurately focus on key historical information. Based on this, the DCS algorithm is introduced to adaptively optimize key hyperparameters of the neural network, significantly improving the model's fitting ability and prediction accuracy. Finally, through the organic combination of quantile regression and kernel density estimation, deterministic point prediction is extended to interval prediction containing complete probability distribution information. Experimental results show that this model, while ensuring high-precision point prediction, can accurately quantify the uncertainty of short-term power load, achieving effective assessment of load fluctuation risk.

Claims

1. A short-term power load probability prediction method based on an improved neural network, characterized in that, Includes the following steps: Step 1: Collect historical power load data and relevant external influencing factor data to determine the input data for the model; Step 2: Preprocess the input data, including outlier handling and normalization, and divide the data into training and test sets; Step 3: Construct a short-term power load probability prediction model based on DCS-CNN-BiLSTM-Attention-QR neural network; the model includes an input layer, a convolutional layer, a bidirectional long short-term memory network layer, an attention mechanism layer, a fully connected layer, a quantile regression layer, and a kernel density estimation layer connected in sequence; the quantile regression layer uses a quantile loss function instead of a traditional loss function; Step 4: Initialize the parameters of the prediction model and set the initial parameters of the differential creation search algorithm; Step 5: Train the model using the training set data, use the root mean square error of point prediction as the fitness function of the DCS algorithm, optimize the number of hidden nodes, regularization coefficient and learning rate of the bidirectional long short-term memory network layer, and use quantile loss as the loss function for model training to obtain the trained probability density prediction model. Step 6: Use the trained model to predict the test set data to obtain the predicted power load values ​​under multiple set quantiles, forming a prediction sequence; Step 7: Based on the predicted sequence, the probability density curve of the prediction interval is generated using the kernel density estimation method to complete the probabilistic prediction of short-term power load.

2. The short-term power load probability prediction method according to claim 1, characterized in that, In step one, the input data includes load data, electricity price, and time type for the same time in the previous seven days.

3. The short-term power load probability prediction method according to claim 1, characterized in that, In step two, the mean square method is used to process outlier data, and the maximum-minimum normalization method is used to map the data to the [0,1] interval; the ratio of training set to test set is 80%:20%.

4. The short-term power load probability prediction method according to claim 1, characterized in that, In step four, the optimization range for the number of hidden layer nodes in the bidirectional long short-term memory network is [0, 100], the optimization range for the regularization coefficient is [1E-6, 1E-2], and the optimization range for the learning rate is [1E-5, 1E-2]. The population size of the DCS algorithm is 30, the maximum number of iterations is 20, and the ratio of high, medium, and low performers in the population is 1:2:

3.

5. The short-term power load probability prediction method according to claim 1, characterized in that, In step six, the quantile regression layer is set with 9 quantiles, the quantile values ​​range from (0,1), and the interval is 0.

1.

6. The short-term power load probability prediction method according to claim 1, characterized in that, In step seven, the kernel density estimation method employs a Gaussian kernel function and adaptive bandwidth.

Images