CN111342471A

CN111342471A - Machine learning-based family obstetrician and consumer power optimization management method

Info

Publication number: CN111342471A
Application number: CN202010135401.3A
Authority: CN
Inventors: 周华嫣然; 周羿宏; 胡俊杰
Original assignee: North China Electric Power University
Current assignee: North China Electric Power University
Priority date: 2020-03-02
Filing date: 2020-03-02
Publication date: 2020-06-26
Anticipated expiration: 2040-03-02
Also published as: CN111342471B

Abstract

The invention discloses a family prosumer power optimization management method based on machine learning, which comprises the following steps of firstly, establishing a family prosumer power management system model based on machine learning; then, collecting historical data, substituting the historical data into a long-term and short-term memory LSTM network for off-line training; and finally, inputting online collected data into the LSTM, generating a scheduling result online, and issuing a scheduling instruction to the stored energy. The method manages the household electricity utilization conditions under the excitation of time-of-use electricity price, including charging and discharging of photovoltaic power, electricity utilization load and stored energy, and the management aims at minimizing the electricity utilization cost. The invention controls the charging and discharging of the stored energy on line according to the conditions of the photovoltaic power generation power and the electric load so as to achieve the management target.

Description

Machine learning-based family obstetrician and consumer power optimization management method

Technical Field

The invention belongs to the field of optimal scheduling of power systems, and particularly relates to a family producer and consumer power optimal management method based on a machine learning technology.

Background

To mitigate environmental pollution and climate change, sustainable/renewable energy sources have been substantially incorporated into the power grid. To make better use of the random renewable energy, more and more households have installed photovoltaic panels and battery energy storage devices to transform them from pure consumers of electric energy to consumers for the production and consumption of electric energy (consumers with power generation capacity). To maximize the benefits of photovoltaic energy storage systems, Home Energy Management Systems (HEMS) are deployed by home producers and consumers to schedule and coordinate the use of electrical energy. Demand Response (DR), which can effectively reduce electricity consumption and costs of residents, has received increasing attention in recent years. Providing electricity price incentives to home owners and consumers to optimize electricity usage profiles is one of the common ways to implement demand response.

Because the household power demand and the photovoltaic power generation are random, the optimization problem borne by the HEMS is a continuous decision process in time with different uncertainty sources. Traditional methods for solving these stochastic problems are Mixed Integer Linear Programming (MILP), etc., but their high online computational burden and strong prediction dependencies make it difficult to integrate their algorithms into computing devices with limited computing power in HEMS.

With the progress of computer computing power in recent years, intelligent computing technologies represented by artificial neural networks, recurrent neural networks and support vector machines have been applied to optimization and computation of various power systems on a large scale by combining big data technologies and applying data driving models such as machine learning. By mining information among mass data, the problem with rapidity and certain accuracy can be solved by applying a machine learning technology, and the calculation burden of online solving is reduced. However, the existing applications mostly focus on load prediction, classification and the like, and the research on the power optimization management of the HEMS is less. Currently, the existing research on the machine learning technology on the HEMS power optimization management mainly adopts two methods: deep learning and reinforcement learning. In the method based on reinforcement learning, although the optimized arrangement of the power consumption of the household appliances and the single electric vehicle has been studied, the state of the equipment is processed under the technique based on reinforcement learning, and actually, in the home consumer system, the state is continuous and complex, and the speed of processing data cannot follow the speed of state change. Deep learning is very good at processing high dimensional data and extracting patterns from it very quickly, and therefore, there is a study on neural network technology that combines deep learning in reinforcement learning to reduce the dimension of the device state. However, because the reinforcement learning process is the system state, the application of the reinforcement learning related method can only control the on-off of the device, the power of the device cannot be continuously adjusted, and the potential of power optimization management is not fully excavated. The power of the equipment can be continuously optimized and adjusted by utilizing the deep learning technology, and the power optimization management method has a better prospect on the aspect of HEMS power optimization management by combining the capability of processing high-dimensional data through deep learning. In the aspect of deep learning application, a plug-and-play algorithm is proposed to solve the problem of power optimization of family producers and consumers on the basis of a recurrent neural network, but the bidirectional recurrent neural network used by the method is not only based on historical data, but also based on future data such as electric forecast and the like, and is still difficult to get rid of prediction. In addition, a new idea is provided for the application of machine learning in the power optimization of household producers and consumers by adopting a clustering mode, but the accuracy of a result is difficult to ensure by a rough clustering mode.

It is considered that, compared with the recurrent neural network, the Long Short-Term Memory (LSTM) neural network not only has the advantages that the recurrent neural network can find the correlation of time sequences and has the capability of memorizing previous states, but also can avoid the problem of gradient disappearance and gradient explosion when the recurrent neural network is trained. The special structure of the LSTM and the design of the forgetting gate in the structure thereof enable the LSTM to have stronger memory capability than a common recurrent neural network, and can show better performance in the time sequence decision problem.

Therefore, the invention adopts the LSTM neural network to carry out power optimization management on the household prosumers and consumers, and utilizes the LSTM neural network to learn the historical optimal solution offline to determine the energy storage power of the household prosumers and consumers online. The long-term state of the system can be considered in the decision making of the family producer and the consumer in the off-line learning process, the HEMS can quickly execute scheduling during on-line operation, the problem of bottom layer optimization during on-line operation is not needed to be solved, the approximate optimal solution is obtained under the condition of not depending on the prediction of the state of the system, and the calculation burden is greatly reduced.

Object of the Invention

The invention aims to perform power optimization management on family producers and consumers by using an LSTM neural network based on a new machine learning technology, and determine the energy storage power of the family producers and consumers on line by using the LSTM neural network to learn a historical optimal solution offline. The long-term state of the system can be considered in the decision making of the family producer and the consumer in the off-line learning process, the HEMS can quickly execute scheduling during on-line operation, the problem of bottom layer optimization during on-line operation is not needed to be solved, the approximate optimal solution is obtained under the condition of not depending on the prediction of the state of the system, and the calculation burden is greatly reduced.

Disclosure of Invention

The invention provides a family lying-in and stillbirth power optimization management method based on machine learning, which comprises the following steps:

a, establishing a family producer and consumer power management system model based on machine learning;

b, collecting historical data, substituting the historical data into a long-term and short-term memory LSTM network for offline training;

and C, inputting online collected data into the LSTM network, generating a scheduling result online, and issuing a scheduling instruction to the stored energy.

Preferably, the process of establishing the machine learning-based home obstetrician and stillboard power management system model in the step a is as follows: firstly, configuring an LSTM network for a family producer and consumer system consisting of a family; then, setting input and output of the LSTM network; for convenience of management, the total scheduling time interval is divided into T time intervals with the length of tau, and power scheduling is performed once for each time interval; setting the output of each LSTM network in the current time period as the value of the state of charge (SOC) of the stored energy in the next time period, wherein the input quantity is the system state quantity needing to be considered for predicting the SOC, and the system state quantity comprises time-of-use electricity price, load electricity power and photovoltaic power generation power; after the input and the output of the network are set, establishing an SOC prediction model of the LSTM network in the next period as shown in the formula (1):

wherein, s (t) represents the SOC of the household energy storage in the t-th time period, namely the output of the LSTM network; n represents the temporal dependence of the prediction, i.e. predictionHow many previous states need to be considered for the energy storage SOC for the next period; u denotes the system state quantity to be taken into account for predicting the SOC, i.e. the input of the LSTM network, which includes the time of use price (c)_TOU) Photovoltaic output power (P)_PV) Electric power (P) for load_d) (ii) a f represents the mapping relationship between the input and output obtained by the LSTM network; the symbol ^ is used for representing a predicted value, namely a decision result of the SOC predicted by the LSTM;

i.e., the predicted SOC of the stored energy during the t +1 th period.

The above equation (1) is written as equation (2) according to the memory capacity of the LSTM network:

wherein k is used to indicate the t + k th predicted state, relating to the t-n +1 th history state to the t + k-1 th history state;

substituting the input variables, the home prosumer power management model may be expressed as equation (3):

after the energy storage SOC predicted in the next period is obtained according to the formula (3), the charge and discharge power of the stored energy is calculated by the formula (4):

wherein, P_B(t) the power of the energy storage battery in the tth time interval, wherein the value of the power is more than 0 to represent charging power, and the value of the power is less than 0 to represent discharging power; c_bIndicating the capacity of the energy storage cell in kWh, tau indicating the specified duration of each time period η_cAnd η_dRespectively, charge efficiency and discharge efficiency.

Further preferably, the collected historical data in the step B is time-of-use electricity price, load electricity power and photovoltaic power generation power data of the family for at least one past year; after the historical data are collected, an optimization model is established by using a linear programming algorithm, and the optimal energy storage charge-discharge power configuration corresponding to each group of historical data is solved by taking the day as a unit; then, configuring an optimal solution by using the historical data and the energy storage charging and discharging corresponding to each group of historical data to form a training data set of the LSTM network; scaling the data in the training data set to be in a range of [0-1] through minimum-maximum standardization, and substituting the data set into an LSTM network for off-line training after the data set is sorted; selecting Adam as an optimizer of the LSTM network, setting the learning rate of the Adam to be 0.01, setting other parameters to be default values, and setting the trained LSTM network to have 1 input layer, 1 LSTM layer, 1 full-connection layer and 1 regression layer, wherein the input layer has 4 nodes, the LSTM layer has 20 nodes, the full-connection layer has 1 node, and the regression layer has 1 node.

Still more preferably, the process of collecting the input data online in step C is: in the online operation stage, when a new time interval begins, the power management system of the household producer and consumer collects the time-of-use electricity price, the photovoltaic power generation power and the load electricity consumption power at the moment, and updates the input vector; after the collection is finished, a scheduling result is generated on line by using an LSTM network; after the input is updated, the power management system of the household producer and consumer operates the LSTM network, predicts the SOC of the stored energy in the next time period, calculates the charge-discharge power of the stored energy according to the formula (4), generates a power control signal and sends the power control signal; the power management system of the family producer and consumer continuously and repeatedly starts the processes of online data acquisition, generating the SOC of the next time interval by using the LSTM neural network, calculating a power control signal by the SOC and controlling the charging and discharging power of the stored energy, thereby finishing the optimal management of the stored energy power.

Drawings

Fig. 1 is a flow chart of the power optimization management method for home stewardess.

FIG. 2 is a comparison graph of SOC scheduling results curves for 5 consecutive days using the linear programming MILP algorithm and the machine learning LSTM algorithm: FIGS. (a) - (e) are from day 1 to day 5, respectively.

Detailed Description

The method for optimizing and managing the power of the family parity victims based on the machine learning is explained in detail in the following by combining the attached drawings.

The family producer and consumer power optimization management method based on machine learning manages the household electricity utilization conditions under the excitation of time-of-use electricity price, including photovoltaic power, electricity utilization load and charging and discharging of stored energy, and the management aims at minimizing the electricity utilization cost. The method needs to control charging and discharging of stored energy on line according to the conditions of photovoltaic power generation power and electric load so as to achieve the management target. Fig. 1 is a flow chart of the power optimization management method for home stewardess.

The preferred management method comprises the following steps:

step A: and establishing a family producer and consumer power management model based on a machine learning technology. The specific process includes the substeps of A1-A4.

Step A1: an LSTM neural network is configured for the power management system. In principle, a family's system of birth and death who is composed of a family configures an LSTM neural network. The LSTM neural network is applied because the power optimization management of home producers and consumers is a time-sequential decision process, and as a recurrent neural network, LSTM can find the correlation between time sequences, so its performance on the time sequence problem is really better than that of the standard feedforward neural network. In addition, with long and short term memory and forgetting gate structures, LSTM has a stronger memory capacity than the normal recurrent neural network, showing good performance in many tasks.

Step A2: the input and output of the LSTM network are set. For management, we divide the total scheduling time interval into T periods of length τ, and perform power scheduling once per period. Since the state of charge (SOC) of the stored energy can reflect the scheduled power level and has strong continuity in time, the output of each LSTM network in the current period is set to the SOC value of the next period. The input quantity is a system state quantity which needs to be considered for predicting the SOC, and comprises time-of-use electricity price, load electricity power and photovoltaic power generation power.

Step A3: and establishing an SOC prediction model of the LSTM network in the next period. Based on steps A1 and A2, a mathematical model of the SOC in the next period is established as shown in equation (1).

Where s (t) represents the SOC of the home energy storage during the t-th period, i.e. the output of the LSTM network. n represents the predicted time dependence, i.e. how many previous states need to be considered to predict the energy storage SOC for the next period. u denotes the system state quantity to be taken into account for predicting the SOC, i.e. the input of the LSTM network, which includes the time of use price (c)_TOU) Photovoltaic output power (P)_PV) Electric power (P) for load_d). f represents the mapping between the input and output obtained by the LSTM network. The symbol (^) is used to represent the predicted value, i.e., the SOC decision result predicted by LSTM. Thus, the prediction of the stored energy in the t +1 th period can be obtained

Considering the memory capability of the LSTM network, equation (1) can be written as equation (2), where k is used to indicate that the t + k th predicted state is related to the t + n +1 th historical state to the t + k-1 th historical state.

Substituting the input variables, the model can be expressed as equation (3).

Step A4: and establishing a current time period scheduling power model based on the LSTM network. Based on the next-period SOC prediction model of the LSTM network established in step a3, after the energy storage SOC predicted in the next period is obtained, the charge and discharge power of the stored energy may be calculated by equation (4).

Wherein, P_BAnd (t) is the power of the energy storage battery in the tth time interval, the value of the power is greater than 0 to represent charging power, and the value of the power is less than 0 to represent discharging power. C_bIndicating the capacity of the energy storage battery in kwh. tau. indicating the specified duration of each period η_cAnd η_dRespectively, charge efficiency and discharge efficiency.

And B: and collecting historical data, and substituting the historical data into an LSTM network for off-line training. The specific process includes the substeps of B1-B4.

Step B1: and collecting mass historical data. And D, historical data needing to be collected are historical values of the state quantity of the optimization system needing to be considered for predicting the SOC in the model built in the step A, namely the time-of-use electricity price, the load electricity power and the photovoltaic power generation power. Because the machine learning technology is a data-driven technology, in order to ensure the application effect of the LSTM network, the possible situations of the system are considered as much as possible, and the contingency is avoided, so that a historical data sample with mass data is needed to have good generalization capability in the face of different actual situations. If the data set is too small, the learned content of the LSTM network may be insufficient to handle the various practical situations, and overfitting problems may also occur. Therefore, we specify that the sample data used should be at least more than the past year.

Step B2: and solving the optimal charge and discharge power of the stored energy from the historical data by using a traditional linear programming optimization algorithm. After the data are collected, an optimization model is established by applying a traditional linear programming algorithm, and the optimal energy storage charge and discharge power arrangement corresponding to each group of historical data is solved by taking the day as a unit. The linear programming optimization algorithm model is shown in (5) to (17):

except that some variables have already been explained in equations (1) to (4), the remaining variables are explained as follows: equation (4) is the scheduling objective of power optimization management, i.e., economic optimization. c. C_FThe price of the power is fixed. P_gFor power exchange between the household and the power grid.

Representing the amount of power flowing from the grid to the home and the power injected into the grid by the home, respectively. Equation (6) is a power balance constraint. The expressions (7) to (10) are intended to ensure

And

at least one value is 0 in the same time period, namely P_gThere can be only one state during a period of time.

And

is an integer variable from 0 to 1.

Andbrespectively the upper and lower limit values of the power flow between the household and the power grid.

(0 or more) and

(less than or equal to 0) respectively represents the power of energy storage charging and power discharging to the power grid,

and

respectively form the charge and discharge limits.

And is also a dimensionless variable 0-1, which functions to ensure that the stored energy has only one state of charge or discharge during a period of time. S^max,S^minRespectively representing the upper limit and the lower limit of the energy storage SOC.

Step B3: and integrating training data sets. And optimally solving the stored energy charging and discharging arrangement corresponding to the historical data and each group of historical data to form a training data set of the LSTM network. Given that the magnitude of the data in the training dataset is different and that the LSTM is sensitive to the magnitude of the data, it is desirable to scale these data to the range of 0-1 by min-max normalization.

Step B4: and training the LSTM network. Substituting into LSTM network for off-line training. To train the LSTM network, we chose Adam as the optimizer for the network because of its computational efficiency and good performance. We set Adam's learning rate to 0.01 and other parameters to default values. On the basis of comparing the prediction accuracy and the network complexity, the trained LSTM network has 1 input layer, 1 LSTM layer, 1 fully-connected layer and 1 regression layer, wherein the input layer has 4 nodes, the LSTM layer has 20 nodes, the fully-connected layer has 1 node, and the regression layer has 1 node.

And C, inputting online collected data into the LSTM, generating a scheduling result online, and issuing a scheduling instruction to the stored energy. The specific process comprises the substeps of C1-C2.

Step C1: the input data is collected online. In the online operation stage, when a new time step begins, the power management system collects the time-of-use electricity price, the photovoltaic power generation power and the load electricity power at the moment, and updates the input vector.

Step C2: and generating a scheduling result on line by using the LSTM network. After the input is updated, the power management system for the household product and consumption runs the LSTM network, the SOC of the stored energy at the next time step is predicted, the charging and discharging power of the stored energy is calculated by the formula (4), and a power control signal is generated and issued. The power management system for home production and consumption continuously and repeatedly starts the online data acquisition, utilizes the LSTM neural network to generate the SOC of the next time period, calculates the power control signal by the SOC, controls the process of energy storage charging and discharging power, and completes the optimal management of energy storage power.

Fig. 2 shows the energy storage SOC scheduling result of a household producer and consumer equipped with a photovoltaic panel and a battery for 5 consecutive days under the excitation of time-of-use electricity price, and the result is respectively solved by using the conventional MILP algorithm (global optimal solution) and the LSTM algorithm of the present invention, and comparing the SOC curve of 5 days shows that the solution result of the LSTM is not much different from the energy storage SOC of the conventional algorithm, the charging and discharging trend is basically consistent, and the solution result of the LSTM has approximate optimality.

ADVANTAGEOUS EFFECTS OF INVENTION

The invention has the following advantages:

1. different from the traditional power optimization management method for household producers and consumers, the method puts the calculation amount of the problem in an off-line stage, does not need to calculate the underlying mathematical problem during on-line operation, only needs to match the input and output relationship through a trained LSTM network, quickly obtains an approximately optimal scheduling signal, has high calculation efficiency, reduces the calculation pressure of HEMS, and has one-time calculation time in millisecond level after multiple times of simulation.

2. Although the method does not need any prediction of the future state quantity of the system when solving, the method can still obtain an approximately optimal solution. The mapping relation between massive historical data and the corresponding optimal solution is learned in the training process of the LSTM network, and the memory of the LSTM network can ensure that the decision result utilizing the LSTM network has certain global property, so that the LSTM network has approximate optimality without depending on prediction.

3. The method has generality in modeling mode, has no special application condition, has wide application range, and is easy to be popularized and applied to various family producer and consumer power optimization scheduling processes.

Claims

1. A family prosumer power optimization management method based on machine learning is characterized in that the prosumer refers to a consumer with power generation capacity, and the method comprises the following steps:

2. The method for optimized power management of home stewardess based on machine learning as claimed in claim 1, wherein the process of establishing the model of the home stewardess power management system based on machine learning in step A comprises:

firstly, configuring an LSTM neural network for a family producer and consumer system consisting of a family;

then, setting the input and output of the LSTM network, dividing the total scheduling time interval into T time intervals with the length of tau for the convenience of management, and performing power scheduling once for each time interval;

setting the output of each LSTM network in the current time period as the value of the state of charge (SOC) of the stored energy in the next time period, wherein the input quantity is the system state quantity needing to be considered for predicting the SOC, and the system state quantity comprises time-of-use electricity price, load electricity power and photovoltaic power generation power; after the input and the output of the network are set, establishing an SOC prediction model of the LSTM network in the next period as shown in the formula (1):

wherein, s (t) represents the SOC of the household energy storage in the t-th time period, namely the output of the LSTM network; n represents the predicted time dependence, i.e. how many previous states need to be considered to predict the energy storage SOC for the next period; u denotes the system state quantity to be taken into account for predicting the SOC, i.e. the input of the LSTM network, which includes the time of use price (c)_TOU) Photovoltaic output power (P)_PV) Electric power (P) for load_d) (ii) a f represents the mapping relationship between the input and output obtained by the LSTM network; the symbol ^ is used for representing a predicted value, namely a decision result of the SOC predicted by the LSTM;

namely the predicted SOC of the stored energy in the t +1 th period;

3. The family producer and consumer power optimization management method based on machine learning as claimed in claim 2, wherein in step B, the collected historical data is time-of-use electricity price, load electricity power, photovoltaic power generation power data of the family at least for more than one past year;

after the historical data are collected, an optimization model is established by using a linear programming algorithm, and the optimal energy storage charge-discharge power configuration corresponding to each group of historical data is solved by taking the day as a unit; then, configuring an optimal solution by using the historical data and the energy storage charging and discharging corresponding to each group of historical data to form a training data set of the LSTM network; scaling the data in the training data set to be in a range of [0-1] through minimum-maximum standardization, and substituting the data set into an LSTM network for off-line training after the data set is sorted; selecting Adam as an optimizer of the LSTM network, setting the learning rate of the Adam to be 0.01, setting other parameters to be default values, and setting the trained LSTM network to have 1 input layer, 1 LSTM layer, 1 full-connection layer and 1 regression layer, wherein the input layer has 4 nodes, the LSTM layer has 20 nodes, the full-connection layer has 1 node, and the regression layer has 1 node.

4. The family indigestion person power optimization management method based on machine learning as claimed in claim 3, wherein the process of online collecting input data in step C is:

in the online operation stage, when a new time interval begins, the power management system of the household producer and consumer collects the time-of-use electricity price, the photovoltaic power generation power and the load electricity consumption power at the moment, and updates the input vector;

after the collection is finished, a dispatching result is generated on line by using an LSTM network, when the input is updated, the power management system of the family producer and consumer operates the LSTM network, the SOC of the stored energy in the next time period is predicted, the charging and discharging power of the stored energy is calculated by the formula (4), and a power control signal is generated and issued;

the power management system of the household producer and consumer continuously and repeatedly starts the processes of online data acquisition, generating the SOC of the next time interval by using the LSTM network, calculating a power control signal by the SOC and controlling the charging and discharging power of the stored energy, thereby finishing the optimal management of the stored energy power.