CN113962454A - LSTM energy consumption prediction method based on dual feature selection + particle swarm optimization - Google Patents

Abstract

The invention discloses an LSTM energy consumption prediction method based on dual feature selection and particle swarm optimization. The method comprises the following steps: performing correlation analysis on time and feature dimensions of the original data set by adopting an MI mutual information method, and selecting front N' dimension features most effective on the energy consumption prediction target value; step two: performing secondary feature selection on the N-dimensional features to obtain N' dimensional features after PMI feature selection; step three: performing model training and prediction on the data after PMI dual feature selection by adopting an LSTM model to obtain an initial prediction sequence y (t); step four: and optimizing the hyperparameter units, dropout and batchsize of the LSTM model by adopting a PSO algorithm, thereby improving the prediction precision of the LSTM model and finally obtaining the PMI-LSTM-PSO model. The method has the advantages of high prediction precision, high algorithm efficiency and stable prediction performance.

Description

LSTM energy consumption prediction method based on dual feature selection + particle swarm optimization

technical field

The invention relates to the technical field of building energy consumption prediction, more specifically, it is an LSTM energy consumption prediction method based on dual feature selection + particle swarm optimization.

Background technique

With the widespread application of more and more complex technological products, the demand for electricity is gradually increasing worldwide, and it is necessary to control the power grid to achieve sustainable development of electricity. In the era of artificial intelligence, the Internet of Things has gradually been integrated into daily life, and the development of smart grids also requires the corresponding testing capabilities, and smart meters emerge as the times require. The continued expansion of smart metering infrastructure across the globe has also laid the foundation for the introduction of active energy systems into the smart grid. The State Grid Corporation of China has been deploying technologies such as smart meters, distribution automation and embedded intelligence on a large scale since launching its "Strong Smart Grid" program in 2009.

For home buildings and enterprise buildings, it is of great practical significance to improve the efficiency of energy consumption and reduce energy consumption through the prediction of energy consumption. Commercial and residential buildings account for 30% to 40% of total energy consumption in smart buildings. Current trends suggest that this percentage is likely to increase in the near future and that energy consumption and penetration are increasing globally. Therefore, short-term energy consumption prediction is very important. Due to the complexity and various uncertainties of the building's infrastructure behavior, as well as the shortcomings of traditional power grids such as low efficiency, serious power waste, weak information interaction ability and low degree of automation, the This becomes a challenging problem.

With this in mind, researchers have developed a number of forecasting methods to improve grid quality and optimize energy use. In many related studies, time series models such as ARIMA are often used as reference models to verify whether some newly proposed methods have superior prediction performance. Currently researchers often use historical data in conjunction with machine learning and deep learning algorithms, such as artificial neural networks (ANN), support vector machines (SVM), adaptive neuro-fuzzy inference systems (ANFIS) and extreme learning machines (ELM) to make predictions . Among them, convolutional neural network and BP neural network have been studied in the field of electricity consumption, but they are still in the initial stage of prediction methods.

In the process of data preprocessing, the quality of feature selection on the original data largely determines the accuracy of the model. Predictive models are better enhanced if the number of input data features can be reduced by selecting the most efficient and useful inputs. The methods of feature selection methods include correlation analysis and numerical sensitivity analysis, but these methods are all linear input selection methods, while energy consumption data are nonlinear. Therefore, the mutual information feature selection method will be more effective, and the efficiency of this method to calculate the correlation between input and output data is very high. Feature variable selection based on mutual information is a novel variable selection method, in which mutual information quantifies and calculates the correlation between different correlated variables.

1) MI mutual information algorithm

Mutual Information (MI) represents the interdependence between two variables X and Y.

The mutual information I(X; Y) between X and Y is defined as:

Among them, p(x, y) is the joint probability density function, p(x), p(y) are the marginal probability density functions of x and y, respectively. MI is used to evaluate the amount of information that the occurrence of one event contributes to the occurrence of another event. The MI mutual information method achieves the purpose of feature selection by calculating the mutual information measure of all features and target features, and then sorting them to select N' features with the highest correlation.

2) Person correlation coefficient

分别为X,Y的平均值。如果r≥0.5说明X,Y之间相关性较强，否则说明X,Y之间相关性较弱。通过Person相关系数进行二次特征选择，可进一步减少特征。in,

are the mean values of X and Y, respectively. If r≥0.5, it means that the correlation between X and Y is strong; otherwise, it means that the correlation between X and Y is weak. Quadratic feature selection by Person correlation coefficient can further reduce features.

3) LSTM model

LSTM is a deep learning model that can efficiently process longer time series and automatically learn data and mine deeper features. However, similar to other neural network models, the setting of some hyperparameters in the LSTM neural network model often relies on the experience of researchers, and such models lack scientific rigor. The advantage of PSO is that it is simple and easy to implement, the PSO solution provides faster convergence, and there are not many parameters to adjust. Genetic algorithm and ant colony algorithm do not have this kind of guidance mechanism.

为当前时刻信息，o_t为输出门，

表示矩阵元素相乘，

表示矩阵相加。Long-short-term neural memory network (LSTM) was proposed by Hochreiter to solve the gradient disappearance and gradient explosion problems of Back-propagation Through Time (BPTT). As the model continues to improve, it gradually evolves into the widely used LSTM network architecture. Its interior is composed of 3 unique gate structures and a state module for storing memory. The structure of the LSTM unit is shown in Figure 1. Among them, C _t is the state information stored by the LSTM unit, h _t is the output of the hidden layer of the unit, f _t is the forget gate, and i _t is the input gate,

is the current moment information, o _t is the output gate,

represents the multiplication of matrix elements,

Represents matrix addition.

Forget Gate: Controls the degree to which the previous unit state C _t-1 is forgotten:

f _t =σ(W _f ·[h _t-1 ,x _t ]+b _f ) (3)

Input Gate: Controls what information is added to this unit:

i _t =σ(W _i ·[h _t-1 ,x _t ]+b _i ) (4)

Cell state update: Selectively record new information into C _t according to f _t :

Output gate: activates C _t and controls how much C _t is filtered:

o _t =σ(W _o ·[h _t-1 ,x _t ]+b _o ) (7)

W_o为各个模块对应的权重矩阵，b_f，b_i，

b_o为偏置项，σ为sigmoid激活函数，tanh为双曲正切激活函数，定义为W _f , W _i ,

W _o is the weight matrix corresponding to each module, b _f , b _i ,

b _o is the bias term, σ is the sigmoid activation function, and tanh is the hyperbolic tangent activation function, which is defined as

σ(x)=1/(1+e ^-x ) (9)

tanh(x)=(e ^x -e ^-x )/(e ^x +e ^-x ) (10)

The output layer passes h _t through a fully connected layer (dense) according to formula (11) to obtain the final predicted value y _t :

Among them, W _y and _by are the weight matrix and the bias term, respectively.

_{y t} =σ(W _y ·h _t +by ) (11)

LSTM controls the transmission of historical information through the gate function, and has certain time series processing and forecasting capabilities.

4) PSO particle swarm optimization algorithm

The basic idea of particle swarm algorithm: a group of birds randomly fly somewhere in a certain area to search for food, all birds only know the distance between themselves and the food and the location information of other birds. When each bird leaves its current location to fly to another location, it will rely on the following information: the surrounding area of the bird that is currently closest to the food, and judge the location of the food based on its own flying experience.

PSO is initialized as a group of random particles (random solution). Then iteratively find the optimal solution. In each iteration, the particle updates itself by tracking two "extremes" (the local optimal solution pbest and the global optimal solution gbest). After finding these two optimal values, the particle updates its velocity and position by the following formula.

v _i =v _i +c ₁ ×rand()×(pbest _i -x _i )+c ₂ ×rand()×(gbest _i -x _i ) (12)

x _i = x _i +v _i

Among them, i=1, 2, ..., N, N is the total number of particles in the particle swarm.

v _i : the current velocity of the ith particle

rand(): random number between (0, 1)

x _i : the current position of the i particle

c ₁ and c ₂ : learning factors

pbest _i and gbest _i are the local optimal position and the global optimal position of the current particle swarm, respectively.

However, the existing MI mutual information algorithm, LSTM model, and PSO particle swarm optimization algorithm have low accuracy for energy consumption prediction, and the prediction performance is unstable, which does not meet the requirements of building energy consumption prediction. Therefore, it is necessary to develop a building energy consumption prediction method with high prediction accuracy and stable prediction performance.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide an LSTM energy consumption prediction method based on multi-dimensional feature selection + particle swarm optimization, which is an energy consumption prediction method applied to buildings, with high prediction accuracy and stable prediction performance.

In order to achieve the above purpose, the technical solution of the present invention is: a method for predicting energy consumption based on MI-LSTM-PSO, which is characterized in that: as shown in FIG. 2 , it includes the following steps:

Step 1: Use the MI mutual information method to analyze the correlation between the time and feature dimensions of the original data set, and select the most effective front N' dimension features for the energy consumption prediction target value, so as to eliminate redundant data and improve the model algorithm. the role of efficiency;

Step 2: Calculate the pearson correlation coefficient value between the first N'-dimensional feature selected by the MI mutual information method and the predicted sequence, and select the N"-dimensional feature whose pearson correlation coefficient value is greater than or equal to 0.5;

Step 3: Use the LSTM model to perform model training and prediction on the N"-dimensional feature data after the PMI double feature selection, and obtain the initial prediction sequence y(t);

Step 4: Use the particle swarm optimization PSO algorithm to optimize the hyperparameters units, dropout, and batchsize of the LSTM model, thereby improving the prediction accuracy of the LSTM model, and finally obtain the MI-LSTM-PSO model.

In the above technical solution, in step 1 and step 2, N' is 60, that is, the first 60-dimensional features that are most effective for the energy consumption prediction target value are selected.

In the above technical solution, step 1 specifically includes the following steps:

S11, use the sliding window to form the 20-dimensional feature data of the first 24 hours into 24M (ie, 480)-dimensional feature components, wherein the original data sequence includes: photovoltaic power generation in 2 regions, energy consumption of different facilities in 17 regions, system grid input Total power (the data sequence will be different according to different scene datasets);

S12, using the MI mutual information method to perform feature selection on the above 24M (ie 480) dimension feature components;

where p(x,y) is the joint probability density function of x and y, and p(x) and p(y) are the marginal density functions. If x is completely uncorrelated with y, then p(x,y) will be equal to p(x)p(y), its mutual information will be equal to 0, if I(X; Y) is larger, it means that the correlation between the two variables is stronger;

S13, determine the optimal parameter N of the MI feature selection dimension through experimental optimization; if the value of N is too large, the model training data set will contain too much redundant information and noise, which will deteriorate the prediction performance, and the value of N is too large. Small, the model training data set contains too little information, which will also make the prediction results worse; usually, the optimal N value is between 3M and 6M, the prediction performance is better, and the N value is relatively small. ;

S14 , based on the mutual information ranking of the feature sequence x(t) and the target sequence Y, synthesizing the time and feature dimension data, select the top 60-dimensional features that are most effective for predicting the target value of energy consumption as the training data set of the subsequent model.

In the above technical solution, step 2 specifically includes the following steps:

S21, calculate the Pearson correlation coefficient between the above 60-dimensional feature components and the target sequence Y (ie Gi);

分别为X,Y的平均值；如果r≥0.5说明X,Y之间相关性较强，否则说明X,Y之间相关性较弱；in,

are the average values of X and Y respectively; if r≥0.5, it means that the correlation between X and Y is strong, otherwise, it means that the correlation between X and Y is weak;

S22, according to the fact that the pearson correlation coefficient is less than 0.5, the correlation between the two is weak, and the 37-dimensional feature data with the pearson correlation coefficient greater than or equal to 0.5 is selected.

In the above technical solution, the LSTM network includes three gate structures and a state module for storing memory. As shown in Figure 1, step 3 specifically includes the following steps:

为当前时刻信息，o_t为输出门，“×”表示矩阵元素相乘，“+”表示相加运算，σ为sigmoid函数；S31, set C _t as the state information stored by the LSTM unit, x _t as the input of the input layer, h _t as the output of the hidden layer of the unit, f _t as the forgetting gate, i _t as the input gate,

is the current moment information, o _t is the output gate, "×" means multiplication of matrix elements, "+" means addition operation, and σ is the sigmoid function;

S32, forget gate: used to control the degree to which the previous unit state C _t-1 is forgotten, its expression is as follows:

f _t =σ(W _f *[h _t-1 ,x _t ]+b _f ) (3)

S33, input gate: used to control which information is added to this unit, and its expression is as follows:

i _t =σ(W _i *[h _t-1 ,x _t ]+b _i ) (4)

S34, the state information stored by the unit: it is used to selectively record new information into C _t according to _{f t} and it, and its expression is as follows:

S35, output gate: used to activate C _t and control the degree to which C _t is filtered, its expression is as follows:

o _t =σ(W _o *[h _t-1 ,x _t ]+b _o ) (7)

h _t =o _t *tanh(C _t ) (8)

W_o分别为f_t、i_t、

o_t对应的权重矩阵，b_f、b_i、

b_o分别为f_t、i_t、

o_t对应的偏置项，tanh为双曲正切激活函数，定义如下：Among them, h _t is the output of the hidden layer of the unit; h _t-1 is the output of the hidden layer of the previous unit; W _f , _Wi ,

W _o are _ft , _it ,

Weight matrix corresponding to o _t , b _f , b _i ,

b _o are _ft , _it ,

The bias term corresponding to o _t , tanh is the hyperbolic tangent activation function, which is defined as follows:

σ(x)=1/(1+e ^-x ) (9)

tanh(x)=(e ^x -e ^-x )/(e ^x +e ^-x ) (10)

S36, the output layer passes h _t through a fully connected layer to obtain the final predicted value y _t according to the following formula:

_{y t} =σ(W _y *h _t +by ) (11)

In the above formula, W _y and _by are the weight matrix and the bias term, respectively.

In the above technical solution, step 4 specifically includes the following steps:

S41, initialize the modification parameters, and set the ranges of the following parameters units∈[20,300], dropout∈[0,1], batchsize∈[20,300];

S42, within the initial range, randomly initialize the particle swarm (20 particles), calculate the fitness value (mean absolute error MAE) of each particle according to the fitness function (LSTM model fitting result), and calculate the fitness value (mean absolute error MAE) of each particle according to the current prediction of each particle The index MAE determines the optimal position of the particle swarm in this iteration (pbest) and the best orientation of the historical particle swarm (gbest);

S43, according to the position and velocity of the optimal particle, update the position and velocity of the current particle, fit the updated particle through the LSTM model, calculate the MAE of each particle, and update pbest and gbest according to the MAE;

v _i =v _i +c ₁ ×rand()×(pbest _i -x _i )+c ₂ ×rand()×(gbest _i -x _i ) (12)

x _i = x _i +v _i

In formula (12): i = 1, 2, ..., N, N is the total number of particles in the particle swarm;

v _i : the current velocity of the i-th particle;

rand(): a random number between (0, 1);

x _i : the current position of the i particle;

c ₁ and c ₂ : learning factors;

pbest _i and gbest _i are the local optimal position and the global optimal position of the current particle swarm, respectively;

S44, after the updated particles are trained by the LSTM model, the fitness value of each particle is calculated, and the optimal position of the particle swarm in this iteration and the optimal orientation of the historical particle swarm are updated according to the fitness value;

S45, when the fitness value of the optimal particle no longer changes or the number of iterations reaches the upper limit value, it is considered that the algorithm has reached convergence; if the particle does not converge, continue to return to S33 to update the particle;

S46: Substitute the obtained optimal particle parameters units, dropout, and batchsize into the LSTM model, perform model prediction on the data in step 1, and obtain a final prediction result.

The above "*" all mean: multiply.

The present invention has the following advantages:

(1) The present invention is an energy consumption prediction method applied to buildings, with high prediction accuracy and stable prediction performance;

(2) The present invention reduces 87.5% of redundant features through MI, which plays a good role in improving the efficiency of the model algorithm, and the model algorithm has high efficiency;

(3) The present invention uses the PSO algorithm to optimize the hyperparameters units, dropout and batchsize of the LSTM model, thereby improving the accuracy of the LSTM model prediction, and the model fitting effect is good;

(4) The predicted value of the PMI-PSO-LSTM model in the present invention is basically within the confidence interval of the real value, and the predicted trend is close to the real value, and the prediction accuracy is high;

(5) The MAE and SMAPE of the PMI-PSO-LSTM combined model in the present invention are superior to all results of other models, and have higher robustness and more stable prediction performance.

Description of drawings

Figure 1 is a schematic diagram of the internal structure of the existing LSTM.

FIG. 2 is a schematic structural diagram of the PMI-PSO-LSTM model of the present invention.

FIG. 3 is a graph showing a comparison of prediction results of a basic model according to an embodiment of the present invention.

FIG. 4 is a comparison scatter diagram of the prediction results of the basic model according to the embodiment of the present invention.

FIG. 5 is a graph showing a comparison of prediction results of a combined model according to an embodiment of the present invention.

FIG. 6 is a comparison scatter diagram of the prediction results of the combined model according to the embodiment of the present invention.

FIG. 7 is a comparison diagram of model evaluation indicators according to an embodiment of the present invention.

Detailed ways

The implementation of the present invention will be described in detail below with reference to the accompanying drawings, but they do not constitute a limitation of the present invention, but are merely examples. At the same time, the advantages of the present invention are made clearer and easier to understand by the description.

Example

Now, the present invention will be described in detail by taking the power consumption prediction of a certain building as an example, which also has a guiding role for the present invention to be applied to the energy consumption prediction of other buildings.

In this implementation, the historical electricity consumption of a building is used as a time series for short-term single-step 1h electricity consumption forecast.

In this embodiment, the electricity consumption forecast of a building includes the following contents:

1. Experimental dataset and MI feature selection

The data set used in this example is the electricity consumption of a building from October 15, 2019 to June 4, 2019, and the data set has a total of 20 features. A description of these features is shown in Table 1. The data in the fifth column is the pearsonr correlation coefficient value between the current feature and the Gi feature.

Table 1 Dataset Description

In this embodiment, the data of the previous 24 hours is used to predict the value of Gi in the next hour, so a sliding window is used to form 480 feature components from the data of 20 features in 24 hours. Then the MI mutual information method is used to select the top 60-dimensional features with the largest MI value among the 480 feature components formed using the sliding window method.

The selection results are shown in Table 2; the selection results are shown in Table 2;

Among them, the selected feature, such as Gi(t-1), is input from the public power grid of the industrial plant one hour before the current time as the benchmark;

Table 2 Features selected by MI

Among them, the selected feature, such as Gi(t-1), represents the input from the public grid of the industrial plant one hour before the current time. The MI value is the mutual information value of the current feature component X and the Gi component (that is, I(X; Gi(t)) based on the current time. It can be seen from Table 2 that most of the features in the previous four hours and the Gi feature at the current moment The mutual information value of Gi, Ao, Co, and A2 is relatively large, and the mutual information value of the Gi, Ao, Co, and A2 features in the first 24 hours and the Gi feature at the current moment are also relatively large. MI reduces the redundant features by 87.5%, which is effective in improving the efficiency of the model algorithm. to a good effect.

The data set used in this embodiment is a 20-dimensional feature, and the data of the previous 24 hours is used to predict the data of the 25th hour in the future.

Experiments have been carried out on the data set of 20-dimensional features in this embodiment, and the experimental results show that:

1) The prediction result obtained by selecting the first 60-dimensional features is similar to that of 100-dimensional features;

2) After the dimension of feature data is increased (that is, the dimension of feature data is selected to be greater than 100), the prediction result will be worse;

3) After the dimension of the feature data is reduced (that is, the dimension of the feature data is selected to be less than 60), the data set contains too little information, which will also make the prediction result worse.

Therefore, in this embodiment, the MI mutual information method is used to select the top 60-dimensional features with the largest MI value among the 480 feature components formed by the sliding window method.

2. Evaluation indicators

Use 4 evaluation indicators to judge the quality of the model.

Root mean square error: RMSE, the smaller the value, the better the model fitting effect.

Mean absolute error: MAE, the smaller the value, the better the model fitting effect.

Symmetric mean absolute percentage error: SMAPE, the smaller the value, the better the model fitting effect.

Coefficient of determination: R2, the larger the value, the better the model fitting effect.

为预测值，y_i为真实值，

为真实值的均值，n为数据数量。In formulas (13), (14), (15), (16),

is the predicted value, y _i is the actual value,

is the mean of the true values, and n is the number of data.

3. Model parameter settings

In order to verify the prediction effect of the proposed MI+PSO-LSTM combined model, two groups of 6 experimental models (ie M1-M6) in Table 3 are used for experimental comparison in this embodiment. The main parameters of the models are shown in Table 4 and Table 5. .

Table 3 Experiments vs. benchmark models

No Model describe M1 ARIMA Difference Integrated Moving Average Autoregressive Model M2 KNR K-nearest neighbor (regression) model M3 LSTM LSTM model M4 MI-LSTM Mutual information method + LSTM model M5 PMI-LSTM Mutual information method + LSTM model M6 PMI-LSTM-PSO Mutual information method + PSO optimization LSTM model

Table 4 The main parameters of the comparison model 1

Table 5 The main parameters of the comparison model 2

4. Model experiment data analysis

4.1. Analysis of the experimental results of the basic model

In this embodiment, the basic models M1-M3 in Table 3 are used, and a single-step prediction experiment comparison is performed for the total electricity Gi input to the public power grid through features 1-20.

In the experimental comparison results (Table 6), it can be seen from the four model prediction evaluation indicators of coefficient of determination, root mean square error, and symmetric mean absolute percentage error that the LSTM model has the best prediction results.

Table 6 Comparison of basic model experiments

Model R2 RMSE MAE SMAPE ARIMA 0.872609 12.1688 7.496174 8.548175 KNR 0.849556 13.21612 8.155453 9.543262 LSTM 0.889503 11.211024 6.622012 7.594866

Figures 3 and 4 show the comparison between the predicted results of ARMA, K-nearest neighbors and LSTM to predict 1h electricity consumption and the actual value. It can be seen from Figure 3 and Figure 4 that the trend predicted by the LSTM model is the closest to the real value, and only the LSTM model is in the confidence interval of the original value. The result curve predicted by ARIMA and K-nearest neighbor model is not within the confidence interval of the true value, but also has the problem of prediction lag. In summary, compared with the ARMA and K nearest neighbor regression models, the prediction effect of the LSTM model is the best. So choose LSTM as the experimental base model.

4.2. Analysis of experimental results of LSTM combined model

In this embodiment, the combined models M3-M6 in Table 3 are used, and 20 groups of single-step prediction and comparison experiments are carried out for the total electricity Gi input to the public power grid through features 1-20.

Figure 5 and Figure 6 show the comparison between the prediction results of the four models to predict the electricity consumption Gi for 1 hour and the actual value. It can be seen from Figure 5 and Figure 6 that the predicted values of the four models are basically within the confidence interval of the true value, and the trend predicted by the PMI-PSO-LSTM model is the closest to the true value. It can be seen from FIG. 7 that the evaluation indicators of the PMI-PSO-LSTM model are all optimal (in FIG. 7, M3, M4, M5, and M6 respectively use the combined models M3-M6 in Table 3 in this embodiment).

Table 7 shows the average of the 20 experimental results of the four combined models. The first four columns are the four evaluation indicators of the prediction model, and the fifth column is the training time of the prediction model. As can be seen from Table 7, compared with the LSTM, MI-LSTM and PMI-LSTM models, the MI+PSO-LSTM model is not significantly improved on R2, but the performance on MAE and SMAPE is improved by 20% and 10% respectively. , 5% or so. Compared with the LSTM model, the performance of MI-LSTM is not significantly improved, but after selecting features through MI, the dimension of the input data is reduced by 87.5%, resulting in a reduction of model training time by about 63%. Compared with the MI-LSTM model, the performance of the PMI-LSTM has not improved almost, but after the secondary feature selection, the dimension of the input data is reduced by about 40%, resulting in a reduction of the model training time by about 20%;

Table 7 Comparison of evaluation indicators of combined models

Model R2 RMSE MAE SMAPE t LSTM 0.88724 11.18282 7.19766 8.56986 159S MI-LSTM 0.89722 10.67590 6.66639 7.82360 59S PMI-LSTM 0.92301 10.73256 6.42070 7.49299 46S MI-PSO-LSTM 0.90482 10.27717 6.12843 6.87869 44S

Figure 7 is a boxplot of the four evaluation indicators of the 20 groups of M3-M6 experiments. The '+' symbols that are not within the box shape in Figure 7 are outliers (can be ignored). As can be seen from Figure 7, the four evaluation indicators of the MI-PSO-LSTM model are significantly better than the other three models, and the MAE and SMAPE of the MI-PSO-LSTM model each time are better than all the results of the other models, while The MI-PSO-LSTM model also outperforms other models by about 95% of the data in R2 and RMSE each time. The four evaluation indicators of MI+LSTM and LSTM partially overlap, but the overall trend of MI-LSTM is better than the LSTM model. As can be seen from Figure 7, compared with the LSTM model, the MI-LSTM model and the PMI-LSTM model, the MI-PSO-LSTM model has the smallest boxplot shape (the difference between the upper and lower quartiles), which shows that the MI-PSO - LSTM model is more stable than other models.

To sum up, the present invention proposes a short-term energy consumption combined prediction model based on PMI, PSO and LSTM. First, in the data preprocessing stage, double feature selection is performed on the original data using the mutual information method and the Pearson correlation coefficient to remove redundant features. Then use PSO to match and optimize the network architecture of LSTM, so that the topological structure of LSTM is best suited to the current input data. Finally, the data after feature selection is input into the optimized LSTM, and the energy consumption data is analyzed. short-term forecast. In order to verify the effect of the MI-PSO-LSTM model on short-term energy consumption prediction, the present invention conducts a multi-dimensional single-step prediction comparison experiment on the energy consumption time series data set of a building. The results of the above experiments show that the four evaluation indicators of the MI-PSO-LSTM combined model are all optimal, which means that the MI-PSO-LSTM model has higher prediction accuracy, robustness and more stable prediction performance. The MI-PSO-LSTM combined model can provide a useful research idea for exploring the predictive analysis aspects of time series using deep learning. However, the MI-PSO-LSTM combined model still has a lot of room for optimization, such as studying the noise filtering problem of time series and the problem of dynamic intelligent selection of features, so as to further optimize the prediction accuracy of the model.

Other unexplained parts belong to the prior art.

Claims (5)

1. A LSTM energy consumption prediction method based on dual feature selection and particle swarm optimization is characterized by comprising the following steps: comprises the following steps of (a) carrying out,

the method comprises the following steps: performing correlation analysis on time and feature dimensions of the original data set by adopting an MI mutual information method, and selecting front N' dimension features most effective on the energy consumption prediction target value;

step two: performing secondary feature selection on the N-dimensional features selected in the step one by adopting a Person correlation coefficient to obtain N' dimensional features after PMI feature selection;

step three: carrying out model training and prediction on the N' dimensional feature data after PMI feature selection by adopting an LSTM model to obtain an initial prediction sequence y (t);

step four: and optimizing the hyperparameter units, dropout and batchsize of the LSTM model by adopting a particle swarm optimization PSO algorithm, thereby improving the prediction precision of the LSTM model and finally obtaining the PMI-LSTM-PSO model.

2. The LSTM energy consumption prediction method based on dual feature selection + particle swarm optimization according to claim 1, wherein: the first step specifically comprises the following steps of,

s11, forming the M-dimensional feature data of the first 24 hours into 24M-dimensional feature components using a sliding window, wherein the original data sequence includes: photovoltaic power generation of 2 areas, energy consumption of 17 different facilities of the areas, and total electric quantity input by a system power grid;

s12, selecting the characteristics of the 24M dimensional characteristic components by using an MI mutual information method;

in formula (1): p (X, Y) is a joint probability density function of X and Y, and p (X) and p (Y) are marginal density functions, if X and Y are not related at all, p (X, Y) will be equal to p (X) p (Y), and mutual information will be equal to 0, if I (X; Y) is larger, the correlation between the two variables is stronger;

s13, determining the optimal parameter N of MI feature selection dimension through experimental optimization; if the value of N is too large, the model training data set will contain too much redundant information and noise, which will deteriorate the prediction performance, while if the value of N is too small, the model training data set will contain too little information, which will also deteriorate the prediction result; generally, the optimal N value is between 3M and 6M, and the feature dimension with better prediction performance and smaller N value is selected;

and S14, based on mutual information sequencing of the characteristic sequence x (t) and the target sequence Y, integrating time and characteristic dimension data, and selecting the first N' dimension characteristic most effective on the energy consumption prediction target value as a training data set of a subsequent model.

3. The LSTM energy consumption prediction method based on dual feature selection + particle swarm optimization according to claim 2, wherein: the second step specifically comprises the following steps:

s21, calculating a Pearson correlation coefficient of the N' -dimensional characteristic component and the target sequence Y;

in formula (2):

respectively the average values of X and Y;

if r is more than or equal to 0.5, the correlation between X and Y is stronger, otherwise, the correlation between X and Y is weaker;

s22, selecting the N' dimension characteristic data with pearson correlation coefficient larger than or equal to 0.5.

4. The LSTM energy consumption prediction method based on dual feature selection + particle swarm optimization according to claim 3, wherein: the LSTM network internally comprises three gate structures and a state module for storing and memorizing, and the third step specifically comprises the following steps:

s31, setting C_tFor the state information stored for the local LSTM cell, x_tAs input to the input layer, h_tFor the output of the hidden layer of this unit, f_tTo forget the door, i_tIn order to input the information into the gate,

as information of the current time o_tFor the output gate, "×" indicates matrix element multiplication, "+" indicates addition operation, σ is sigmoid function;

s32, forget gate: for controlling the last cell state C_t-1The degree of forgetting, the expression of which is as follows:

f_t＝σ(W_f*[h_t-1,x_t]+b_f) (3)

s33, input gate: for controlling which information is added to the unit, the expression is as follows:

i_t＝σ(W_i*[h_t-1,x_t]+b_i) (4)

s34, cell stored state information: for according to f_tAnd i_tSelectively recording new information to C_tWherein the expression is as follows:

s35, output gate: for mixing C_tActivating and controlling C_tThe degree of filtering is expressed as follows:

o_t＝σ(W_o*[h_t-1,x_t]+b_o) (7)

h_t＝o_t*tanh(C_t) (8)

formula (3) to formula (8): w is_f、W_i、

W_oAre respectively f_t、i_t、

o_tCorresponding weight matrix, b_f、b_i、

b_oAre respectively f_t、i_t、

o_tThe corresponding bias term, tanh, is a hyperbolic tangent activation function, defined as follows:

σ(x)＝1/(1+e^-x) (9)

tanh(x)＝(e^x-e^-x)/(e^x+e^-x) (10)

s36, the output layer is h_tObtaining the final predicted value y through a full connection layer_t：

y_t＝σ(W_y*h_t+b_y) (11)

In formula (11): w_yAnd b_yRespectively, a weight matrix and an offset term.

5. The LSTM energy consumption prediction method based on dual feature selection + particle swarm optimization according to claim 4, wherein: the fourth step specifically comprises the following steps of,

s41, initializing modification parameters, setting the range units belonging to [20,300], dropout belonging to [0,1], batchsize belonging to [20,300 ];

s42, randomly initializing the particle swarm in an initial range, calculating an adaptive value of each particle according to the fixness function, and determining pbest of the iterated particle swarm and gbest of the historical particle swarm according to the prediction index MAE of each current particle;

s43, updating the position and the speed of the current particle according to the position and the speed of the optimal particle, fitting the updated particle through an LSTM model, calculating the MAE of each particle, and updating pbest and gbest according to the MAE;

v_i＝v_i+c₁×rand()×(pbest_i-x_i)+c₂×rand()×(gbest_i-x_i) (12)

x_i＝x_i+v_i

in formula (12): i is 1, 2, …, N is the total number of particles in the population;

v_i: the current velocity of the ith particle;

and rand (): a random number between (0, 1);

x_i: i current position of the particle;

c₁and c₂: a learning factor;

pbest_iand gbest_iRespectively obtaining a local optimal position and a global optimal position of the current particle swarm;

s44, after the updated particles are trained through an LSTM model, calculating the adaptive value of each particle, and updating the optimal position of the particle swarm of the iteration and the optimal orientation of the historical particle swarm according to the adaptive value;

s45, when the fitness value of the optimal particle is not changed any more or the iteration number reaches the upper limit value, the algorithm is considered to have converged at the moment; if the particle is not converged, the flow returns to S33 to update the particle;

and S46, substituting the obtained optimal particle parameters units, dropout and batchsize into the LSTM model, and performing model prediction on the data in the first step to obtain a final prediction result.

Images