Energy management method of photovoltaic energy storage system based on probability prediction

Abstract

A probabilistic forecasting-based energy management method for photovoltaic energy storage systems uses a long short-term memory (LSTM) neural network to obtain photovoltaic probability forecasts. This method then samples multivariate distributions based on copula function models to generate forecast scenarios. Based on the forecast results, model predictive control is used to provide real-time optimization, aiming to maximize benefits and minimize costs while meeting load requirements and energy storage characteristics. This method quantifies forecast uncertainty, reduces the robustness optimization losses associated with traditional point forecasting, improves the safety and reliability of the energy storage system, and achieves efficient utilization of photovoltaic energy.

Description

Technical Field

[0001] The present invention belongs to the technical field of photovoltaic power generation, and in particular relates to an energy management method for a photovoltaic energy storage system based on probability prediction. Background Art

[0002] Solar energy boasts advantages such as large energy resources, easily developed resources, and clean, pollution-free operation. Therefore, photovoltaic power generation is considered an important means of resolving the energy crisis. However, the random and volatile nature of photovoltaic power generation leads to irregular and random output variations, which introduces uncertainty into power system planning and operation. This is especially true when photovoltaic penetration is high. Therefore, energy storage units such as batteries are needed to smooth fluctuations and store energy for future use. Large-scale photovoltaic deployment requires rapid and accurate forecasting of on-site short- and long-term photovoltaic power generation output to effectively manage energy demand, energy storage, and the supply of supplementary and backup energy sources.

[0003] For estimating and predicting PV maximum power, we primarily use data-driven methods, including statistical techniques and machine learning tools. We estimate PV power based on environmental information, such as array irradiance, module temperature, ambient temperature, and wind speed, as well as historical data. This allows us to construct a prediction model that correlates PV power output with environmental information, including irradiance, ambient temperature, and wind speed.

[0004] Energy management methods primarily include forecasting, day-ahead optimization, and real-time optimization. Energy management is performed based on forecast results, and real-time optimization is performed. Energy storage, acting as an energy buffer, can meet fluctuating solar radiation and load demands. While utilizing time-of-use electricity pricing, it can save money by storing energy at low-cost, off-peak rates. Furthermore, it can use stored energy to reduce peak demand and generate revenue, while improving the controllability, stability, and reliability of the entire power system.

[0005] The current energy management methods for photovoltaic energy storage systems to solve photovoltaic uncertainty and absorption problems have the following drawbacks:

[0006] 1. Data-driven PV output forecasting methods significantly degrade in performance under low irradiance and regression coefficients. Due to the seasonal nature of light conditions, the forecast model requires regular updates to maintain accuracy. Furthermore, deterministic forecasts based on the conditional mean point of the forecast signal provide very limited information for energy management decisions, and forecast deviations introduce significant uncertainty.

[0007] 2. Most energy management approaches typically only consider the energy state of energy storage, ignoring the system's dynamic characteristics, such as charging and discharging rate limitations. They also ignore the lifetime cost of energy storage systems when considering time-of-use electricity pricing benefits. Rule-based short-term real-time optimization calculations lack the potential impact on subsequent strategy selection. Summary of the Invention

[0008] In view of the technical problems existing in the background technology, the energy management method of photovoltaic energy storage system based on probability prediction provided by the present invention uses a long short-term memory (LSTM) neural network to obtain photovoltaic probability prediction, and samples multivariate distribution based on the copula function model method to generate prediction scenarios. Based on the prediction results, while meeting the load demand and energy storage characteristics, with the goal of maximizing benefits and minimizing costs, model predictive control (MPC) is used to provide real-time optimization, quantify the uncertainty of the prediction to reduce the robustness optimization loss caused by traditional point prediction, improve the safety and reliability of the energy storage system, and realize the efficient consumption and utilization of photovoltaic energy.

[0009] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0010] A photovoltaic energy storage system energy management method based on probability prediction includes the following steps:

[0011] A photovoltaic energy storage system energy management method based on probability prediction is characterized by comprising the following steps:

[0012] S1 generates photovoltaic prediction scenarios and uses an LSTM neural network to obtain probabilistic predictions, including multiple data such as temperature, wind speed, cloud cover, cloud type, and humidity as input variables, while also adding historical information on solar radiation values ​​as input variables; the solar radiation value is used as the output;

[0013] S2, normalize all input variables, unify the dimensions, remove the units, and input them according to time periodicity;

[0014] Define a statistical parametric model of the prediction error to assume the distribution of uncertainty, and use an LSTM neural network to predict the parameter-specified distribution.

[0015] For the prediction results (solar irradiance), solve the statistical parameters of the prediction results, such as mean, variance, skewness, and kurtosis;

[0016] S3, takes the historical data of all input variables as the input of prediction, divides 70% of the historical data set into training set and 30% into validation set, and adds weight noise during training to ensure the noise information in the data, sets the network type and prediction range , the number of hidden layers and neurons;

[0017] The training set is used to complete the training. To estimate the statistical quality of the point forecast, the degree of correspondence between the forecast and the actual observation is calculated. The root mean square error is used as the error metric and the specified quantile of the predicted target distribution is output. ;

[0018] S4, using the copula function model-based method to model the marginal distribution of each input variable, collect historical data of the input variables, and transform each variable into a single variable by using the probability integral transformation on the marginal distribution of the single variable. The data points are converted into D cubes , thereby estimating the copula density;

[0019] After obtaining the multivariate distribution, the dependency structure of the copula model and the original data is obtained:

[0020] ,

[0021] Use the copula model to generate multivariate random vectors in two steps to generate dependent random numbers from a uniform distribution Sampling in , Indicates the first generated number in a group, and the others u2, u3, etc. are generated (or solved) from the conditional distribution function; according to the conditional distribution function Generate in sequence ;

[0022] In the formula, U and u have different meanings. Uniform distribution U is a fixed mathematical term, and u is a custom data u i A collection of

[0023] S5, based on the marginal distribution from the probability prediction, the generated variables of unit D cube are converted into the original variable dimension using inverse transformation sampling to obtain the predicted photovoltaic power generation scenario and the corresponding quantile;

[0024] S6, based on photovoltaic forecasts, time-of-use electricity price charts and load change charts, and based on the MPC algorithm, performs energy management of the photovoltaic energy storage system;

[0025] Determine the output power based on a fixed PV system size With irradiance included The relationship between the input variables included;

[0026] The energy storage system state model includes batteries and converters, and calculates energy conversion, state of charge, and battery cycle loss costs. and state of charge SOC, based on time-of-use electricity prices Calculating the cost of charging and discharging the grid ,According to the dynamic characteristics of the energy storage system, a state space model is established;

[0027] S7, establish system constraints, including SOC upper and lower limits, battery charge and discharge current limits, and the initial SOC and the final SOC being equal or approximately equal;

[0028] Determine the objective functionJ Expressed as:

[0029]

[0030] Where, , are the weight coefficients for point prediction and 50% probability prediction respectively; C BE is the battery loss cost, C grids is the cost of charging and discharging the grid;

[0031] S8, set the time-of-use electricity price and energy storage equipment parameters, the target initial SOC value, select the MPC step size, prediction steps, and control steps for simulation, solve the objective function, and output the control variables ; Complete control operations based on the prediction results.

[0032] Preferably, in step S2, photovoltaic output prediction is performed using 10 minutes as the time scale input variable.

[0033] Preferably, in step S3, the prediction range is set to 24 hours. .

[0034] Preferably, in step S7, the system constraint is modified by the soft constraint as follows:

[0035]

[0036] In the formula, SOC represents the state of charge, SOC min Indicates the minimum state of charge, P BE Indicates the battery output (discharge) power, P dis_max Indicates the maximum discharge power: P ch_max Indicates the maximum charging power.

[0037] Preferably, in step S7, the function weight coefficient is increased according to the energy management target requirements. Improve expected returns and reduce system conservatism costs.

[0038] Preferably, in step S8, the MPC step sizes are selected as 15 minutes, 30 minutes and 1 hour respectively according to the predicted duration, load variation and calculation requirements.

[0039] Preferably, in step S6, when the size of the photovoltaic system is determined, its output power and irradiance The relationship is:

[0040]

[0041] Where, To predict photovoltaic power generation, is the photovoltaic panel conversion efficiency, is the conversion efficiency affected by temperature, Indicates the output power under nominal conditions. represents the predicted irradiance, Indicates the nominal irradiance;

[0042] The energy storage system state model includes batteries and converters, and its energy conversion is:

[0043]

[0044] In the formula is the output power of the energy storage system, To exchange power between the photovoltaic energy storage system and the grid, Consumes power for the load;

[0045] The state of charge of the energy storage system is calculated as:

[0046]

[0047] Where, is the state of charge of the battery at time k+1, is the energy conversion efficiency of the battery, Indicates the rated power of the battery under nominal conditions;

[0048] Energy storage systems based on lithium batteries, battery cycle loss cost Calculated as:

[0049]

[0050] Where, is the total purchase cost of the battery, The number of effective cycles of the battery throughout its life cycle;

[0051] Based on time-of-use electricity prices The cost of charging and discharging the grid Calculated as:

[0052]

[0053] According to the dynamic characteristics of the energy storage system, the state space model is established as:

[0054]

[0055] In the formula, SOC is selected as the state variable ,choose As a control variable .

[0056] This patent can achieve the following beneficial effects:

[0057] 1. This invention uses neural networks for probabilistic prediction and copula sampling to generate prediction scenarios, effectively reducing the uncertainty of photovoltaic predictions and quantifying the uncertainty of predictions to reduce the loss of traditional point prediction errors to decision optimization.

[0058] 2. The present invention utilizes MPC controller simulation, based on the system's dynamic characteristics modeling, to perform predictive control under the constraints of charge and discharge rate and state of charge, limiting the maximum voltage and current output, maintaining the energy storage system's SOC within a safe range, and ensuring system safety;

[0059] 3. The energy management method of the present invention fully considers the impact of photovoltaic output fluctuations, time-of-use electricity prices, and energy storage system losses. Based on the MPC method of multi-step prediction and single-step control, it improves the overall benefits of the system and increases the utilization rate of photovoltaic power. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] The present invention will be further described below with reference to the accompanying drawings and examples:

[0061] Figure 1 This is a schematic structural diagram of a grid-connected photovoltaic energy storage system according to an embodiment of the present invention;

[0062] Figure 2 A flowchart for generating a photovoltaic prediction scenario in an embodiment of the present invention;

[0063] Figure 3 Flowchart of MPC real-time optimization in an embodiment of the present invention;

[0064] Figure 4 This is a time-of-use electricity price change diagram in an embodiment of the present invention;

[0065] Figure 5 This is a graph showing changes in predicted and actual photovoltaic power generation in an embodiment of the present invention;

[0066] Figure 6 This is a real-time load change diagram in an embodiment of the present invention. DETAILED DESCRIPTION

[0067] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0068] like Figure 1 As shown, for a grid-connected photovoltaic energy storage system, it includes photovoltaic power generation components and a DC-DC converter with photovoltaic MPPT, an energy storage system and a bidirectional DC-DC converter with an MPC control algorithm, a DC bus, AC and DC loads, and a grid-connected inverter;

[0069] The PV panels obtain maximum power through the MPPT algorithm, adjust the output voltage to the reference value, and provide power through the DC-DC converter. The AC and DC loads obtain power through the converter connected to the DC bus. The energy storage system, based on the energy management system MPC algorithm, provides power to the load side through the DC-DC converter, or absorbs power from the PV panel or the grid-connected mains.

[0070] The present invention proposes a photovoltaic energy storage system energy management method based on probability prediction, which includes the following steps:

[0071] S1, such as Figure 2 As shown, PV forecast scenarios are generated, probabilistic forecasts are obtained using LSTM neural networks, and copula models are used for full sampling;

[0072] Select the temperature, wind speed, cloud cover, cloud type, and humidity of the forecast location as input variables, and add the historical information of solar radiation values ​​as input variables; the solar radiation value is used as the output;

[0073] S2, normalize all variables, unify the dimensions, and remove the units. All input variables have time periodicity and are input at 10-minute intervals.

[0074] Define a statistical parameter model of the prediction error to assume the distribution of uncertainty, and use the LSTM neural network to predict the parameter specified distribution. For the prediction results, solve the statistical parameters of the prediction results;

[0075] S3 uses the historical data of all input variables as the input for prediction, divides 70% of the historical data set into a training set and 30% into a validation set, and adds weight noise during training to ensure the noise information in the data.

[0076] Set the network type, prediction range m, number of hidden layers and number of neurons. In this embodiment, a bidirectional propagation neural network is selected, and m=144, indicating that the prediction range is the next 24 hours.

[0077] Complete training using the training set;

[0078] To estimate the statistical quality of the point forecast, calculate how well the forecast corresponds to the actual observation. The root mean square error (RMSE) is used as the error metric:

[0079] (1)

[0080] Where, is the number of sample data, is the output of the prediction model, is the actual measured value;

[0081] Train the neural network and update the number of hidden layers and neurons based on the root mean square error;

[0082] In this case, the best prediction model is set to 4 hidden layers and 10 neurons;

[0083] Outputs the specified quantile of the predicted target distribution ;

[0084] (2)

[0085] when , this is an estimate of the conditional median of the output distribution.

[0086] S4, such as Figure 3 As shown in Figure 2, sampling a multivariate distribution using a copula-based method consists of two substeps:

[0087] Appropriately model the marginal distribution of each variable and construct a copula model that includes the dependency structure of the multivariate distribution.

[0088] First, we need to collect historical data of variables and transform each variable into The data points are converted into D cubes , thereby estimating the copula density and completing the model.

[0089] After obtaining the multivariate distribution, the dependency structure of the copula model and the original data is obtained:

[0090] (3)

[0091] Use the copula model to generate multivariate random vectors in two steps to generate dependent random numbers from a uniform distribution Sampling in . According to the conditional distribution function Generate in sequence ;

[0092] In S5, based on the marginal distribution from the probability prediction, the generated variables of unit D cube are converted into the original variable dimension using inverse transformation sampling to obtain the predicted photovoltaic power generation scenario and the corresponding quantile.

[0093] S6, such as Figures 4-6 As shown, based on photovoltaic forecasts, time-of-use electricity price charts, load change charts, etc., the photovoltaic energy storage system energy management is performed based on the MPC algorithm;

[0094] When the size of the photovoltaic system is determined, its output power and irradiance The relationship is:

[0095] (4)

[0096] Where, To predict photovoltaic power generation, is the photovoltaic panel conversion efficiency, is the conversion efficiency affected by temperature, Indicates the output power under nominal conditions. represents the predicted irradiance, Indicates the nominal irradiance;

[0097] The energy storage system state model includes batteries and converters, and its energy conversion is:

[0098] (5)

[0099] In the formula is the output power of the energy storage system, To exchange power between the photovoltaic energy storage system and the grid, Consumes power for the load;

[0100] The state of charge of the energy storage system is calculated as:

[0101] (6)

[0102] Where, is the state of charge of the battery at time k+1, is the energy conversion efficiency of the battery, Indicates the rated power of the battery under nominal conditions;

[0103] Energy storage systems based on lithium batteries, battery cycle loss cost Calculated as:

[0104] (7)

[0105] Where, is the total purchase cost of the battery, The number of effective cycles of the battery throughout its life cycle;

[0106] Based on time-of-use electricity prices The cost of charging and discharging the grid Calculated as:

[0107] (8)

[0108] According to the dynamic characteristics of the energy storage system, the state space model is established as:

[0109] (9)

[0110] In the formula, SOC is selected as the state variable ,choose As a control variable ;

[0111] S7: Establish system constraints, including upper and lower SOC limits, battery charge and discharge current limits, and ensure that the initial and final SOC are roughly equal. In this case, to protect the battery from overcharge and overdischarge, the upper and lower SOC limits are 20% and 80%, respectively.

[0112] (10)

[0113] Determine the objective function J Expressed as:

[0114] (11)

[0115] Where, , are the weight coefficients for point prediction and 50% probability prediction, respectively; in this case, they are 1 and 0.1 respectively;

[0116] Select the MPC step size, prediction step number, and control step number based on the duration of TOU electricity price changes and load changes. In this case, the step size is set to 1 hour, the control step number is 1, and the prediction step number is 8.

[0117] Set the time-of-use electricity price and energy storage equipment parameters. In this case, the time-of-use electricity price is as follows Figure 4 As shown, the unit price of the energy storage device is 4 yuan / AH, including 24 groups of 1kWH batteries, and the cycle life is 10,000 times;

[0118] S8, set the target initial SOC value, perform simulation, solve the objective function, and output the control variable ; Complete control operations based on the prediction results.

[0119] The above embodiments are merely preferred technical solutions of the present invention and should not be construed as limiting the present invention. The scope of protection of the present invention shall be the technical solutions set forth in the claims, including equivalent alternatives to the technical features of the technical solutions set forth in the claims. In other words, equivalent alternatives and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A photovoltaic energy storage system energy management method based on probability prediction, characterized in that The following steps are involved: S1 generates photovoltaic prediction scenarios and uses an LSTM neural network to obtain probabilistic predictions. It selects multiple data points at the prediction location, including temperature, wind speed, cloud cover, cloud type, and humidity, as input variables. It also adds historical information on solar radiation values ​​as input variables. The solar radiation value is used as output. S2, normalize all input variables, unify the dimensions, remove the units, and input them according to time periodicity; Define a statistical parametric model of the prediction error to assume the distribution of uncertainty and use an LSTM neural network to predict the parameter-specified distribution; For the prediction results, solve the statistical parameters of the prediction results; S3, takes the historical data of all input variables as the input of prediction, divides 70% of the historical data set into training set and 30% into validation set, and adds weight noise during training to ensure the noise information in the data, sets the network type and prediction range , the number of hidden layers and neurons; The training set is used to complete the training. To estimate the statistical quality of the point forecast, the degree of correspondence between the forecast and the actual observation is calculated. The root mean square error is used as the error metric and the specified quantile of the predicted target distribution is output. ; S4, using the copula function model-based method to model the marginal distribution of each input variable, collect historical data of the input variables, and transform each variable into a single variable by using the probability integral transformation on the marginal distribution of the single variable. The data points are converted into D cubes , thereby estimating the copula density; After obtaining the multivariate distribution, the dependency structure of the copula model and the original data is obtained: ， Use the copula model to generate multivariate random vectors in two steps to generate dependent random numbers from a uniform distribution Sampling in , Represents a set of u, which is the generated dependent random number; according to the conditional distribution function Generate in sequence ; S5, based on the marginal distribution from the probability prediction, the generated variables of unit D cube are converted into the original variable dimension using inverse transformation sampling to obtain the predicted photovoltaic power generation scenario and the corresponding quantile; S6, based on photovoltaic forecasts, time-of-use electricity price charts and load change charts, and based on the MPC algorithm, performs energy management of the photovoltaic energy storage system; Determine the output power based on a fixed PV system size and irradiance relation; The energy storage system state model includes batteries and converters, and calculates energy conversion, state of charge, and battery cycle loss costs. and state of charge SOC, based on time-of-use electricity prices Calculating the cost of charging and discharging the grid ,According to the dynamic characteristics of the energy storage system, a state space model is established; S7, establish system constraints, including SOC upper and lower limits, battery charge and discharge current limits, and the initial SOC and the final SOC being equal or approximately equal; Determine the objective function J Expressed as: ； Where, , are the weight coefficients for point prediction and 50% probability prediction respectively; C BE is the battery loss cost, C grids is the cost of charging and discharging the grid; S8, set the time-of-use electricity price and energy storage equipment parameters, the target initial SOC value, select the MPC step size, prediction steps, and control steps for simulation, solve the objective function, and output the control variables ; Complete control operations based on the prediction results.

2. The method for energy management of a photovoltaic energy storage system based on probability prediction according to claim 1, characterized in that: In step S2, the photovoltaic output is predicted by inputting variables with a time scale of 10 minutes.

3. The method for energy management of a photovoltaic energy storage system based on probability prediction according to claim 1, characterized in that: In step S3, the prediction range is set to 24 hours. .

4. The method for energy management of a photovoltaic energy storage system based on probability prediction according to claim 1, characterized in that: In step S7, the system constraints are modified as follows after the soft constraints are modified: ； In the formula, SOC represents the state of charge, SOC min Indicates the minimum state of charge, P BE Indicates the battery output power, P dis_max Indicates the maximum discharge power: P ch_max Indicates the maximum charging power.

5. The method for energy management of a photovoltaic energy storage system based on probability prediction according to claim 1, characterized in that: In step S7, the function weight coefficient is increased according to the energy management target. Improve expected returns and reduce system conservatism costs.

6. The photovoltaic energy storage system energy management method based on probability prediction according to claim 1, characterized in that: In step S8, the MPC step sizes are selected as 15 minutes, 30 minutes and 1 hour respectively according to the predicted duration, load variation and calculation requirements.

7. The method for energy management of a photovoltaic energy storage system based on probability prediction according to claim 1, characterized in that: In step S6, when the size of the photovoltaic system is determined, its output power and irradiance The relationship is: ； Where, To predict photovoltaic power generation, is the photovoltaic panel conversion efficiency, is the conversion efficiency affected by temperature, Indicates the output power under nominal conditions. represents the predicted irradiance, Indicates the nominal irradiance; The energy storage system state model includes batteries and converters, and its energy conversion is: ； In the formula is the output power of the energy storage system, To exchange power between the photovoltaic energy storage system and the grid, Consumes power for the load; The state of charge of the energy storage system is calculated as: ； Where, is the state of charge of the battery at time k+1, is the energy conversion efficiency of the battery, Indicates the rated power of the battery under nominal conditions; Energy storage systems based on lithium batteries, battery cycle loss cost Calculated as: ； Where, is the total purchase cost of the battery, The number of effective cycles of the battery throughout its life cycle; Based on time-of-use electricity prices The cost of charging and discharging the grid Calculated as: ； According to the dynamic characteristics of the energy storage system, the state space model is established as: ； In the formula, SOC is selected as the state variable ,choose As a control variable .

Images