CN115564123A - Interval forecasting of short-term power load based on multi-objective and Bayesian optimization - Google Patents

Abstract

The method is used for solving the problem that a power supply system after new energy grid connection has high intermittence and volatility, and uncertain information in load data cannot be provided through traditional point prediction. Therefore, the invention provides a short-term power load interval prediction method based on multi-Objective and Bayesian Optimization (MOBO), and belongs to the field of power load prediction. The method comprises the steps of firstly, establishing a quantile regression model based on deep learning, and objectively verifying the processing capacity of each neural network on nonlinear and time sequence information in a load; secondly, analyzing the prediction effects of the single model and the mixed model through validity check, and selecting a model with a reliable prediction interval; and finally, introducing a multi-objective and Bayesian optimization theory to the selected model to construct a deep learning interval prediction model based on the MOBO. The experimental results show that: the proposed model enables a more accurate description of the fluctuation range of the power load.

Description

Interval forecasting of short-term power load based on multi-objective and Bayesian optimization

technical field

The invention belongs to the field of short-term electric load forecasting, and relates to a short-term electric load interval forecasting method based on multiple objective and Bayesian Optimization (MOBO).

Background technique

In my country, the carbon emissions of the power industry account for more than 1/3 of the country's carbon emissions. "Clearly stipulates: build a modern energy system, continue to strengthen the construction of new energy power consumption and cross-regional transmission capacity, orderly promote the centralized development of wind power and photovoltaic power generation, and actively promote the construction of clean energy bases that complement each other. Under the background of this major national strategy, the use and development of renewable energy for power generation has become an important way for decarbonization plans, such as wind power, photovoltaic power and hydropower, etc., and new energy power generation systems will be connected to the grid on a large scale. Affected by climate change and geographical location, the obtained electric energy has great differences and intermittences, which will bring great uncertainty to the production and dispatch of power companies, thus affecting the maintenance of power balance in the grid. Therefore, accurate power load forecasting plays a crucial role in capturing uncertainty, and provides an important basis for grid dispatchers to customize power generation plans and power plant quotations.

At present, the research on power load forecasting is mainly carried out around point forecasting. These forecasting methods are mainly divided into statistical methods (regression analysis method, time series method and gray prediction method, etc.) and artificial intelligence technology (support vector machine, artificial neural network method and expert system approach). However, point forecasting cannot predict and evaluate the possible fluctuation range of forecast results, and it is difficult to meet the high requirements for load forecasting to better capture random information after new energy is connected to the grid; interval forecasting can essentially solve this problem, that is, to Under different confidence levels, the uncertainty in the data is constrained within a controllable range. At present, there are certain studies on interval forecasting at home and abroad. The interval forecasting methods of load mainly include Gaussian process regression, quantile regression, lower limit and upper limit estimation methods, and gray interval forecasting. Among them, the probability prediction method based on quantile regression is flexible and efficient, and can be combined with simple linear models or machine learning.

In addition, the setting and combination of hyperparameters have a significant impact on the predictive performance of machine learning models, such as model accuracy, stability, and generalization ability. Therefore, how to optimize hyperparameters has become one of the most important components in machine learning, and plays a key role in suppressing underfitting and overfitting. Different from hyperparameter optimization methods such as manual search that lacks theoretical support, grid search that is easy to fall into local optimum, and random search that is limited by dimensionality, the Bayesian Optimization (BO) algorithm makes full use of the existing evaluation information To predict the next best hyperparameter combination, which greatly reduces the time required to train hyperparameter combinations, becoming one of the most popular hyperparameter tuning strategies. In addition, most research on interval prediction tends to choose a single objective function as the loss function of hyperparameter training, such as empirical selection methods, root mean square error, and Pinball loss functions, etc., which cannot take into account that the prediction interval covers more real values and the interval average Multiple goals such as smaller width and higher interval precision. The improvement of one goal is often at the expense of other goals. For example, increasing the interval coverage will increase the average width of the interval, and reducing the average width of the interval will also lead to a decrease in the coverage of the interval. Currently, few studies consider the construction of prediction intervals as a multi-objective problem with multiple objectives. Therefore, the introduction of multi-objective and Bayesian optimization theory is of great significance for establishing reliable prediction intervals.

Contents of the invention

In view of this, the object of the present invention is to provide a deep learning interval prediction model based on MOBO. Using the advantages of Bayesian optimization algorithm in the process of hyperparameter training, the convergence speed is fast and the scalability is strong, to find a set of Pareto optimal solutions that make the multi-objective loss function achieve the desired effect, and finally make the constructed prediction interval effective It can accurately capture the uncertainty information in the load data and describe the fluctuation range of future load efficiently.

To achieve the above object, the present invention provides the following technical solutions:

A short-term power load interval forecasting method based on MOBO, specifically comprising the following steps:

S1: Data prediction processing stage. Preprocess the collected raw power load data set, and then divide it into a training set and a test set according to the ratio of 8:2, including for hyperparameter adjustment and interval prediction;

S2: Model building stage. Establish quantile regression models based on deep learning, including models such as CNNQR, RBFQR, LSTMQR, GRUQR, CNN-LSTMQR and CNN-GRUQR;

S3: Prediction interval construction stage. Input the preprocessed data set into the prediction model, and output the prediction interval of the future load at different quantile points according to the confidence interval construction theory;

S4: Model screening stage. Firstly, the models with invalid prediction intervals are preliminarily excluded through the coverage of the prediction intervals, and then the reliability and sophistication of the preliminarily screened prediction models are further evaluated through the Winkler coefficient and the average width of the prediction intervals, and the models with reliable prediction intervals are selected. Model;

S5: MOBO-based hyperparameter optimization stage. Taking into account multiple contradictory interval prediction indicators, a multi-objective function is established as the loss function of the Bayesian optimization algorithm for hyperparameter optimization, and then the best hyperparameter combination is output.

S6: Prediction and evaluation stage. Input the optimized hyperparameters into the selected model to predict the fluctuation range of the future power load, and compare the pros and cons of the proposed model with other models through point prediction and interval prediction evaluation indicators.

Further, in step S1, the sliding window method and the normalization method are used to preprocess the original power load data set, and the time series data is converted into matrix data X _t as the input of the forecasting model.

来更新，表达如下：Further, in step S2, the weight w _τ and bias b _τ of the quantile regression model based on deep learning are initialized at the quantile point τ, and according to the quantile regression theory, the loss function is minimized

To update, the expression is as follows:

The updated matrix weight w _τ ^* and bias b _τ ^* are input as the parameters of the prediction model, and the number of neurons in the hidden layer is set to M, then at time t, input X _t to get the output vector of the hidden layer H ₁ ,H ₂ ,...,H _M , take it as the input of the fully connected layer, and then get the output value at the quantile point τ:

服从(u(X_t),s²(X_t))的高斯分布，u(X_t)是测试样本t的平均分位数,σ²(X_t)则为分位数方差。在给定置信水平100(1-α)％下，预测区间

的定义如下:Further, in step S3, the quantiles at different quantile points τ obtained by formula 3

Obey the Gaussian distribution of (u(X _t ), s ² (X _t )), u(X _t ) is the average quantile of the test sample t, and σ ² (X _t ) is the quantile variance. At a given confidence level of 100(1-α)%, the prediction interval

is defined as follows:

和

分别为预测区间的下、上限，z_1-α/2是标准高斯分布的临界值，取决于置信水平100(1-α)％。In the formula:

and

are the lower and upper bounds of the prediction interval respectively, and z _1-α/2 is the critical value of the standard Gaussian distribution, which depends on the confidence level 100(1-α)%.

Further, in step S5, the short-term power load interval forecasting model based on multi-objective and Bayesian optimization is established, and a multi-objective function is established as The loss function for Bayesian optimization hyperparameters, namely

m ^* = arg min F(m) (7)

Among them, the function F(m) can be equivalently expressed as:

F(m)＝f _mape (m)+f _rmse (m)-f _picp (m)+f _pinaw (m)+f _ws (m) (8)

In the formula, m ^* ＝[m ₁ ,m ₂ ,m ₃ ,m ₄ ] is the Pareto optimal solution obtained by minimizing the multi-objective function, that is, the combination of hyperparameters that brings the greatest benefit, followed by the largest training rounds, the initial Learning rate, learning rate decay period, and learning rate decay rate.

The beneficial effects of the present invention are:

(1) The present invention verifies the processing ability of each neural network on nonlinear and time series data through real power load data sets, and selects the prediction model with the best performance through ablation experiments.

(2) The present invention introduces multi-objective and Bayesian optimization theory, and tunes the hyperparameters, biases and weights of the prediction model. The experimental results prove that the present invention can find parameters suitable for the target task with higher efficiency, and at the same time , which makes up for the phenomenon that the construction of the prediction interval is regarded as a single-objective problem, resulting in extremely high coverage of the prediction interval accompanied by a wide interval width.

Other advantages, objects and features of the present invention will be set forth in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the investigation and research below, or can be obtained from It is taught in the practice of the present invention. The objects and other advantages of the invention may be realized and attained by the following specification.

Description of drawings

In order to make the purpose of the present invention, technical solutions and advantages clearer, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

Fig. 1 is the flow chart of the present invention based on multi-objective and Bayesian optimized short-term power load interval prediction;

Figures 2 and 3 are the results of short-term power load interval forecasting based on multi-objective and Bayesian optimization;

detailed description

Embodiments of the present invention are described below through specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific implementation modes, and various modifications or changes can be made to the details in this specification based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the diagrams provided in the following embodiments are only schematically illustrating the basic concept of the present invention, and the following embodiments and the features in the embodiments can be combined with each other in the case of no conflict.

Please refer to Figure 1, the present invention provides a short-term power load interval measurement method based on multi-objective and Bayesian optimization, which can effectively capture random and uncertain information in load data, and provide guidance for power grid staff in production and scheduling The information for decision-making can make the power system run economically and safely. First, preprocess the collected original power load data set, and then divide it into a training set and a test set according to the ratio of 8:2; then, establish a quantile regression prediction model based on deep learning, and use the confidence interval construction theory to obtain The predicted value of future load at different quantile points; finally, the model with the best prediction performance is screened out through point prediction and interval prediction evaluation indicators; then multi-objective and Bayesian optimization theory is introduced to adjust the hyperparameters of the selected model , using the test set to compare the pros and cons of the proposed model, which specifically includes the following steps:

Step 1: Data prediction processing stage. Carry out sliding window and normalization processing on the collected original power load data set, and then divide it into training set and test set according to the ratio of 8:2, including for hyperparameter adjustment and interval prediction;

Step 2: Model building phase. In order to select a reasonable prediction model, CNN, RBF, LSTM, GRU, CNN-LSTM and CNN-GRU with similar network structures are used for comparative research, and combined with quantile regression theory. Establish a quantile regression model based on deep learning;

Step 3: Prediction interval construction stage. Input the experimental data set into the above prediction model, and construct the theory according to the confidence interval to obtain the prediction interval of the future load at different quantile points;

Step 4: First calculate the coverage of the prediction interval, preliminarily exclude the models whose prediction performance is not up to standard, then further analyze the average width of the prediction interval and Winkler coefficient of the selected model, and then select the model with a reliable prediction interval;

Step 5: MOBO-based hyperparameter optimization stage. Taking into account indicators such as MAPE, RMSE, PICP, PIANW, and WS, establish a multi-objective function as the loss function of the Bayesian optimization algorithm for hyperparameter optimization, and then output the best hyperparameter combination;

Step 6: Prediction stage. Input the optimized hyperparameters into the selected model, and predict the fluctuation range of the future power load, and evaluate the advantages and disadvantages of the prediction model of the present invention and other models through MAPE, RMSE, PICP, PINAW and WS indicators.

The prediction model of the present invention and the PICP index comparison result of other models of table 1

It can be seen from Table 1 that under the four confidence levels, 1) the PICP values in the six models of LSTMQR, GRUQR, CNN-LSTMQR, CNN-GRUQR, MOBO-GRUQR and MOBO-LSTMQR are greater than the given confidence level. Therefore, the above The model can effectively quantify the uncertainty information brought by the power load. 2) In contrast, the PICP values of the RBFQR and CNNQR models are all lower than the given confidence level at four confidence levels. Therefore, in further experiments, RBFQR and CNNQR models are excluded.

Table 2 Comparison results of PINAW and WS indicators between the proposed model and other models

After preliminary screening, the selected model is further evaluated by PINAW and WS indicators. These two indicators can evaluate the reliability and sophistication of the prediction interval. The results are shown in Table 2. The following conclusions can be drawn: Compared with the two mixed The interval prediction effect of the models CNN-LSTMQR and CNN-GRUQR, the other two individual models LSTMQR and GRUQR are better, and their PINAW and WS indicators are lower at the four confidence levels. Therefore, Bayesian and multi-objective optimization will be applied to GRUQR and LSTMQR to output the prediction interval of future load, and the results are shown in Figure 2 and Figure 3.

Finally, it is noted that the above embodiments are only used to illustrate the technical solutions of the present invention without limitation. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be carried out Modifications or equivalent replacements, without departing from the spirit and scope of the technical solution, should be included in the scope of the claims of the present invention.

Claims (5)

1. A short-term power load interval forecasting method based on multiple objectives and Bayesian Optimization (Multiple Objective and BayesianOptimization, MOBO), is characterized in that, the method may further comprise the steps: S1: Data prediction processing stage. Preprocess the collected raw power load data set, and then divide it into training set and test set in proportion, including for hyperparameter adjustment and interval prediction; S2: Model building stage. In order to enhance the reliability and contrast of the experiment, a quantile regression model based on deep learning was established, including a single model and a mixed model; S3: Prediction interval creation stage. Input the preprocessed training set into the prediction model, and output the prediction interval of the future load at different quantile points according to the confidence interval construction theory; S4: Model screening stage. Through the evaluation index, evaluate the reliability and sophistication of the interval prediction of the model, and select the model with an effective prediction interval; S5: MOBO-based hyperparameter optimization stage. Taking into account multiple contradictory prediction interval indicators, establish a multi-objective function as the loss function of the Bayesian optimization algorithm for hyperparameter optimization, and then output the hyperparameter combination with the best prediction effect; S6: Prediction and evaluation stage. Input the optimized hyperparameters into the selected model to predict the fluctuation range of the future power load, and compare the pros and cons of the proposed model with other models through point prediction and interval prediction evaluation indicators. 2. A short-term power load interval forecasting method according to claim 1, characterized in that: in step S1, the original power load data set is preprocessed by using the sliding window method and the normalization method, and the time series data is converted into is the matrix data X _t , which is used as the input of the prediction model. 3. A kind of short-term power load interval prediction method according to claim 1, characterized in that: in step S2, the weight w _τ and the bias b _τ of the quantile regression model based on deep learning are initialized in the fractional The position τ is completed, and according to the quantile regression theory, the loss function is minimized

To update, the expression is as follows:

The updated matrix weight w _τ ^* and bias b _τ ^* are input as the parameters of the prediction model, and the number of neurons in the hidden layer is set to M, then at time t, input X _t to get the output vector of the hidden layer H ₁ ,H ₂ ,...,H _M , take it as the input of the fully connected layer, and then get the output value at the quantile point τ:

4. A kind of short-term power load interval prediction method according to claim 1, characterized in that: in step S3, the quantile Q _yt (τ|X _t ) at different quantile points τ obtained by formula 3 obeys ( Gaussian distribution of u(X _t ), σ ² (X _t )), u(X _t ) is the mean quantile of the test sample t, and σ ² (X _t ) is the quantile variance. At a given confidence level of 100(1-α)%, the prediction interval

is defined as follows:

In the formula:

and

are the lower and upper bounds of the prediction interval respectively, and z _1-α/2 is the critical value of the standard Gaussian distribution, which depends on the confidence level 100(1-α)%. 5. A method for short-term electric load interval forecasting according to claim 1, characterized in that: in step S5, the established short-term electric load interval forecasting model based on MOBO, by taking into account a plurality of contradictory forecast interval indicators, including MAPE , RMSE, PICP, PINAW, and WS, etc., establish a multi-objective function as the loss function of the Bayesian optimization hyperparameter, namely m ^* = argmin F(m) (7) Among them, the function F(m) can be equivalently expressed as: F(m)＝f _mape (m)+f _rmse (m)-f _picp (m)+f _pinaw (m)+f _ws (m) (8) In the formula, m ^* ＝[m ₁ ,m ₂ ,m ₃ ,m ₄ ] is the Pareto optimal solution obtained by minimizing the multi-objective function, that is, the combination of hyperparameters that brings the greatest benefit, followed by the largest training rounds, the initial Learning rate, learning rate decay period, and learning rate decay rate; then, the optimized m ^* is used as the input of the short-term power load interval forecasting model based on multi-objective and Bayesian optimization, and the bias of the proposed model is updated according to step 3 and weights, output prediction intervals about future loads at different quantile points.

Images