CN113283638A

CN113283638A - Load extreme curve prediction method and system based on fusion model

Info

Publication number: CN113283638A
Application number: CN202110434936.5A
Authority: CN
Inventors: 张士成; 陆春艳; 陶晓峰; 吴少雄; 刘涅煊; 熊霞; 邓良柱; 何旭; 张罗平
Original assignee: NARI Group Corp; Nari Technology Co Ltd; NARI Nanjing Control System Co Ltd
Current assignee: NARI Group Corp; Nari Technology Co Ltd; NARI Nanjing Control System Co Ltd
Priority date: 2021-04-22
Filing date: 2021-04-22
Publication date: 2021-08-20

Abstract

The invention discloses a load extreme curve prediction method based on a fusion model, which comprises the following steps: collecting power consumption data related to load extreme value prediction; acquiring power consumption data characteristics from the power consumption data; and inputting the electricity utilization data characteristics into a pre-trained fusion model, and predicting a load extreme value curve. The fusion model is formed by fusing a plurality of XGboost models, a plurality of SLSTM models and a plurality of INDRNN models. A corresponding system is also disclosed. The XGboost model, the SLSTM model and the INDRNN model are fused for prediction, and the problem of difficult prediction caused by the fact that load data has both a stable time sequence and a non-stable time sequence is effectively solved.

Description

Load extreme curve prediction method and system based on fusion model

Technical Field

The invention relates to a load extreme curve prediction method and system based on a fusion model, and belongs to the field of electric power measurement.

Background

The platform area load extreme curve prediction is to effectively predict the maximum load of the platform area in the next 30 days by utilizing a data mining technology under the condition of fully considering the operation characteristics, capacity increasing decision, natural conditions and social influence of the power utilization information acquisition system; the prediction of the platform load extreme curve is beneficial to strengthening the ordered power utilization management, reasonably arranging the operation mode of a power grid, improving the current situation of heavy overload treatment work, improving the treatment efficiency and effect of heavy overload equipment, providing guarantee for the safe and economic operation of a peak-to-peak summer power system, and having important practical significance and economic and social benefits.

Because the load data of different transformer areas have large difference, the load data of some transformer areas are stable time sequences, such as banks and common residential transformer areas, and the load data of some transformer areas are non-stable time sequences, such as coal-to-electricity transformer areas and photovoltaic transformer areas.

Aiming at load data with both a stable time sequence and a non-stable time sequence, the existing load extreme value curve prediction method has great prediction difficulty.

Disclosure of Invention

The invention provides a load extreme curve prediction method and a system based on a fusion model, which solve the problems disclosed in the background technology.

In order to solve the technical problems, the technical scheme adopted by the invention is as follows:

a load extreme curve prediction method based on a fusion model comprises the following steps:

collecting power consumption data related to load extreme value prediction;

acquiring power consumption data characteristics from the power consumption data;

inputting the power utilization data characteristics into a pre-trained fusion model, and predicting a load extreme curve; the fusion model is formed by fusing a plurality of XGboost models, a plurality of SLSTM models and a plurality of INDRNN models.

And collecting power utilization data related to load extreme value prediction, and preprocessing the power utilization data.

The electricity usage data includes current load data and current weather information.

And acquiring power utilization data characteristics from the power utilization data, and screening the power utilization data characteristics by adopting a correlation analysis method.

The fusion model is formed by fusing a first XGboost model, a second XGboost model, a first SLSRTM model, a second SLSRTM model and an INDRNN model.

The process of processing the electricity consumption data characteristics by the fusion model is that,

the first XGboost model inputs the electricity consumption data characteristics, and the output of the first XGboost model is used as the characteristics and is combined with the electricity consumption data characteristics into a first new characteristic;

inputting the first new characteristic into the second SLSTM model, and taking the output of the second SLSTM model as a first prediction result;

inputting a first new characteristic by the first SLSTM model, and combining the output of the first SLSTM model as the characteristic and the power utilization data characteristic into a second new characteristic;

inputting a second new characteristic into the second XGboost model, and taking the output of the second XGboost model as a second prediction result;

inputting a second new characteristic into the INDRNN model, and taking the output of the INDRNN model as a third prediction result;

and fusing the first prediction result, the second prediction result and the third prediction result to obtain a load extreme value curve.

And performing weighted fusion on the first prediction result, the second prediction result and the third prediction result to obtain a load extreme value curve.

A system for load extremum curve prediction based on a fusion model, comprising:

an acquisition module: collecting power consumption data related to load extreme value prediction;

a feature acquisition module: acquiring the characteristics of the power consumption data according to the power consumption data;

a prediction module: inputting the power utilization data characteristics into a pre-trained fusion model, and predicting a load extreme curve; the fusion model is formed by fusing a plurality of XGboost models, a plurality of SLSTM models and a plurality of INDRNN models.

A computer readable storage medium storing one or more programs, the one or more programs comprising instructions, which when executed by a computing device, cause the computing device to perform a fused model-based load extrema curve prediction method.

A computing device comprising one or more processors, one or more memories, and one or more programs stored in the one or more memories and configured to be executed by the one or more processors, the one or more programs including instructions for performing a fused model-based load extremum curve prediction method.

The invention achieves the following beneficial effects: the XGboost model, the SLSTM model and the INDRNN model are fused for prediction, and the problem of difficult prediction caused by the fact that load data has both a stable time sequence and a non-stable time sequence is effectively solved.

Drawings

FIG. 1 is a flow chart of the method of the present invention;

FIG. 2 is an SLSTM architecture;

FIG. 3 illustrates SLSTM parameters;

FIG. 4 is an INDRNN framework;

FIG. 5 is a block diagram of a fusion model;

FIG. 6 is an overall flow chart of training and prediction.

Detailed Description

The invention is further described below with reference to the accompanying drawings. The following examples are only for illustrating the technical solutions of the present invention more clearly, and the protection scope of the present invention is not limited thereby.

As shown in fig. 1, a load extremum curve prediction method based on a fusion model includes the following steps:

step 1, collecting power utilization data related to load extreme value prediction;

step 2, acquiring power consumption data characteristics from the power consumption data;

step 3, inputting the power utilization data characteristics into a pre-trained fusion model, and predicting a load extreme curve; the fusion model is formed by fusing a plurality of XGboost models, a plurality of SLSTM models and a plurality of INDRNN models.

The method adopts a plurality of models with large differences to construct a fusion model, adopts the fusion model to predict the load extreme value curve, can accurately predict the future load extreme value, is favorable for strengthening the ordered power utilization management, reasonably arranges the power grid operation mode, improves the heavy overload treatment working situation, improves the treatment efficiency and effect of heavy overload equipment, provides guarantee for the safe and economic operation of the peak-to-peak summer power system, and has important practical significance and economic and social benefits.

Step 1 and step 2 are that the characteristics are obtained through the electricity utilization data, and the electricity utilization data processing process is consistent when the characteristics are predicted and when the characteristics are trained; during prediction, the collected power utilization data are current load data and current weather information; and when the training characteristic sample is constructed, the collected power utilization data are historical load data and corresponding historical weather information.

After gathering the power consumption data, can carry out the preliminary treatment to the power consumption data, mainly include:

1) preprocessing the load data of the special and public transformer;

deleting a station area with a plurality of records in the same day of one station area in the original data;

deleting the record with the station zone identification being empty in the original data;

removing a platform area with data collected for three months and less than 80 days in the original data;

removing a station area with the maximum active power in three months in the original data being empty for more than 10 days;

removing a distribution area with average active power being empty for more than 10 days in three months in the original data;

removing a platform area with a load rate of over 10 days in three months in the original data;

deleting the records of the current day of the distribution area with the load rate exceeding 300 in three months in the original data;

deleting a station area with the maximum active power of 0 existing in three months in the original data;

deleting the records of the current day of the distribution room with the load rate of 0 in three months in the original data;

deleting a station area with the maximum active power more than 3 times of rated capacity or operation capacity in three months in the original data;

in three months, a station area with the maximum active power 2 times larger than the operation capacity or contract capacity exists, and the average value of the maximum active power of the station area with abnormal data removed is used for repairing (the record and the null value record are removed);

processing the maximum active power null data: repairing the maximum active power average value of the station area with abnormal data removed (removing the record and the null value record);

and (3) average active power null data processing: repairing the average active power average value of the station area with abnormal data removed (removing null value records);

and (3) processing load rate null data: repairing the load rate average value of the station area with abnormal data removed (removing null value records);

and updating the load rate by using the average active power/the maximum active power according to the record that the load rate is empty or exceeds 100, and reserving 2 decimal places.

2) Preprocessing an extreme value of a power supply unit;

repairing the three-month data with the mean value of the week when the missing value of the load extreme value is less than 10 days, and repairing the three-month data with the mean value of the month if the load extreme value is continuously missing;

if the load extreme value missing value is more than 10 days, deleting the power supply unit;

3) preprocessing weather information;

weather conversion rules: 1: sunny; 2: cloud-like; 3: negative; 4: light rain; 5: showering; 6: middle rain; 7: thunderstorm rain; 8: heavy rain; 9: rainstorm; 98: a null value; 99: other values not in the case of 1, 2, 3, 4, 5, 6, 7, 8, 9, 98;

the wind speed conversion rule is as follows: 1: 1-2 level; 2: a small level 3; 3: the grade 3-4 is less than grade 3; 4: converting from less than 3 grade to 3-4 grade; 5: 3-4 grade; 6: 4-5 grade; 98: a null value; 99: other values than in the case of 1, 2, 3, 4, 5, 6, 98.

And acquiring power utilization data characteristics from the preprocessed power utilization data, and selecting data which possibly influence the load extreme value of the transformer area, such as daily load extreme value air temperature, humidity, wind speed and the like by the characteristics. Taking the construction of the training sample as an example, the load extreme value in the previous 30 days, the temperature, the humidity, the wind speed and the like in the previous 30 days can be selected, and the load curve in the future 30 days is taken as a label.

For model prediction and learning, the more features are, the better, and those features with relatively low correlation with the label adversely affect the model prediction and learning. Therefore, a correlation analysis method can be adopted to carry out certain screening on the characteristics, the characteristics corresponding to the load data can adopt an ACF (active carbon fiber) or PACF (Picture archiving and communication function) method, and the weather information can adopt methods such as Pearson analysis and the like for the characteristics.

The fusion model specifically comprises a first XGboost model, a second XGboost model, a first SLSTM model, a second SLSTM model and an INDRNN model, namely two XGboost models, two SLSTM models and an INDRNN model.

The XGboost model is trained by using input data and labels as parameters to be transmitted, the key point is parameter adjustment, and the following important parameter descriptions are mainly selected:

1. boost: indicating the selected weak learner model, and selecting a gbtree by default;

2. objective: the selection of the target function is carried out because of the linear regression problem, and reg, linear, is selected;

3. n _ estimator (num _ boosting _ rounds): this is the number of maximum trees generated, i.e. the maximum number of iterations;

4. left _ rate: sometimes also called eta, the default value of the system is 0.3. The step length of each step of iteration is important, the operation accuracy is not high when the step length is too large, and the operation speed is slow when the step length is too small. A larger learning space can be provided for the model by controlling the learning rate of each time;

5. max _ depth: the system default is 6. We often use a number between 3 and 10, which is the maximum value of the depth of the XGBoost spanning tree. The size of the max _ depth value determines whether the model learning is concrete. The larger the learning, the more concrete the overfitting problem is caused, so that an appropriate value needs to be selected. Setting to 0 represents no limitation, with a value range of [0, ∞ ];

6. lambda: the system defaults to 0. The regularization of XGboost is influenced by the regulation of a regularization item L2 and lambda parameters of the weight, so that the regulation needs attention to the scale, and the overfitting of the model can be reduced by taking a larger value;

7. alpha: the default value of the system is 0, and the L1 regularization term of the weight can accelerate the algorithm speed under the high-dimensional data characteristics.

Although parameter adjustment can improve the diagnosis accuracy of the model, the improvement is limited, and the purpose of parameter adjustment is mainly to prevent overfitting and improve the training speed. Meanwhile, the GPU version of XGboost is arranged in the official open source community of Python, and the training speed is greatly accelerated by adopting the version.

Aiming at the characteristics of load prediction data, a network architecture and parameters adopting an SLSTM model are shown in figures 2 and 3, the network architecture adopts a 3+1 structure (3 layers of LSTM and 1 layer of full connection layer), a multi-layer architecture can carry out deeper expression on data, a cell is the number of neurons in each layer, and the input and output of the vector dimension represent the network data; taking the first layer input as an example, (none, 12, 1) represents (number of samples, time step, variable dimension), respectively.

INDRNN model: RNN has gradient disappearance and explosion problems, while LSTM/GRU solves the problem to some extent, but because tanh or sigmoid is adopted as the activation function of the saturation, the interlayer gradient attenuation test is serious when multiple layers are stacked. Considering new RNN structures, unsaturated activation functions such as Relu can be used, stacked multilayer use, etc. The problems of gradient disappearance and explosion can be effectively solved by adjusting the gradient back propagation based on time, and the INDRNN is independent weight and decouples the neurons in the RNN layer to make them independent from each other, thereby improving the interpretability of the neurons.

The INDRNN model architecture is shown in fig. 4, where Weight and current + ReLU denote input processing and loop processing of each step with ReLU as an activation function, and by stacking this basic architecture, a deep indnnn network can be constructed.

At present, model fusion is adopted to fuse a plurality of models, and the accuracy of prediction is enhanced. Common fusion methods include weighted averaging and Stacking.

Weighted average: the fusion mode of majority weighted voting generally defaults to the same importance of all weighted voting models. But in general, we may attach more importance to those better-performing weighted voting models, and thus need to assign more voting weight to them. In such a fusion of weighting and voting, a model with poor performance can only enhance the importance and persuasion of the model by fusing with other models and obtaining the same voting result.

And (3) Stacking: it is essentially a layered structure, here for simplicity only the secondary stacking is analyzed, assuming we have 3 basis models M1, M2, M3.

The basis model M1, trained on the training set train, then used to predict the tag columns for train and test, P1, T1, respectively. The same work was repeated for M2 and M3, which also gave P2, T2, P3, T3. P1, P2, P3 and T1, T2, T3 are combined to obtain a new training set and test set train2, test2, respectively. Train2 with model M4 of the second layer, predict test2, get the final label list.

Although the Stacking method can effectively improve the model prediction accuracy, the structure is complex, corresponding rules are difficult to obtain, the method is difficult in engineering application, and overfitting of the model can be caused; the fusion method used by the invention uses the Stacking idea for reference, simplifies the construction rule, and finally, the results are weighted and fused with the three model outputs, thereby further improving the generalization of the models.

The first XGBoost model, the second XGBoost model, the first slss model, the second slss model, and the indrn model are fused as shown in fig. 5, specifically as follows:

A1) the first XGboost model (namely XGboost1 in the figure) inputs the electricity data characteristics, and the output of the first XGboost model is used as the characteristics and is combined with the electricity data characteristics into a first new characteristic;

A2) inputting the first new feature into the second SLSTM model (i.e. SLSTM2 in the figure), and outputting the second SLSTM model as the first prediction result;

A3) inputting a first new feature by the first SLSTM model (i.e. SLSTM1 in the figure), and combining the output of the first SLSTM model as the feature and the electricity consumption data feature into a second new feature;

A4) inputting a second new feature into a second XGboost model (namely XGboost2 in the figure), and taking the output of the second XGboost model as a second prediction result;

A5) inputting a second new characteristic into the INDRNN model, and taking the output of the INDRNN model as a third prediction result;

A6) performing weighted fusion on the first prediction result, the second prediction result and the third prediction result to obtain a load extreme value curve; wherein, the initial weighting proportion of each result is set as 0.3, 0.3 and 0.4, and the optimal weighting proportion parameter is found out again by a grid search method.

As shown in fig. 6, before prediction, a training feature sample is constructed by using historical data and the like, then a fusion model is trained, and the model can be used after training is completed, that is: and predicting by using the trained fusion model according to the characteristics of the current data, and storing the prediction result into a database so as to facilitate query. And adding the predicted result and the corresponding characteristics into a training characteristic sample to construct a new training characteristic sample, and periodically training the fusion model by adopting the new training characteristic sample to further improve the accuracy of model prediction.

The method integrates an XGboost model, an SLSTM model and an INDRNN model to form a multilayer structure, extracts the stable time sequence characteristics by the XGboost and puts the stable time sequence characteristics into the SLSTM and the INDRNN for training, extracts the non-stable time sequence characteristics by the SLSTM and puts the non-stable time sequence characteristics into the XGboost and the INDRNN for training, and finally synthesizes the results of the three models for output, thereby effectively solving the problem of difficult prediction caused by the fact that load data has both stable time sequences and non-stable time sequences.

a feature acquisition module: acquiring power consumption data characteristics from the power consumption data;

As will be appreciated by one skilled in the art, embodiments of the present application may be provided as a method, system, or computer program product. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.

The present application is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the application. It will be understood that each flow and/or block of the flow diagrams and/or block diagrams, and combinations of flows and/or blocks in the flow diagrams and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

The present invention is not limited to the above embodiments, and any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention are included in the scope of the claims of the present invention which are filed as the application.

Claims

1. A load extreme curve prediction method based on a fusion model is characterized by comprising the following steps:

collecting power consumption data related to load extreme value prediction;

2. The method according to claim 1, wherein the load extremal curve prediction method based on the fusion model comprises: and collecting power utilization data related to load extreme value prediction, and preprocessing the power utilization data.

3. The method according to claim 1 or 2, wherein the load extremum curve prediction method based on the fusion model comprises: the electricity usage data includes current load data and current weather information.

4. The method according to claim 1, wherein the load extremal curve prediction method based on the fusion model comprises: and acquiring power utilization data characteristics from the power utilization data, and screening the power utilization data characteristics by adopting a correlation analysis method.

5. The method according to claim 1, wherein the load extremal curve prediction method based on the fusion model comprises: the fusion model is formed by fusing a first XGboost model, a second XGboost model, a first SLSRTM model, a second SLSRTM model and an INDRNN model.

6. The method according to claim 5, wherein the load extremal curve prediction method based on the fusion model comprises: the process of processing the electricity consumption data characteristics by the fusion model,

7. The method of claim 6, wherein the load extremum curve prediction method based on the fusion model comprises: and performing weighted fusion on the first prediction result, the second prediction result and the third prediction result to obtain a load extreme value curve.

8. A system for load extremum curve prediction based on a fusion model, comprising:

9. A computer readable storage medium storing one or more programs, characterized in that: the one or more programs include instructions that, when executed by a computing device, cause the computing device to perform any of the methods of claims 1-7.

10. A computing device, comprising:

one or more processors, one or more memories, and one or more programs stored in the one or more memories and configured to be executed by the one or more processors, the one or more programs including instructions for performing any of the methods of claims 1-7.