Multi-mode household storage intelligent scheduling method and system based on large model prediction

Abstract

The invention discloses a multi-modal household storage intelligent scheduling method and system based on large model prediction, and relates to the technical field of energy storage, and the method comprises the steps: obtaining multi-modal data in real time, and carrying out the preprocessing operation, the multi-modal data comprising electricity price, load electricity demand, photovoltaic power generation and weather data; inputting the preprocessed multi-modal data into the large model, and generating prediction results of each preset time period in a future period, including prediction results of electricity price, load electricity demand and photovoltaic generating capacity; based on a prediction result of the large model, adopting a multi-objective optimization algorithm to generate an initial charging and discharging scheduling strategy of each preset time period; and applying the initial charging and discharging scheduling strategy to the household storage system, and continuously optimizing the large model according to an execution effect so as to adjust the charging and discharging scheduling strategy in a future preset time period in a rolling manner. Multi-modal input variables are predicted in real time by introducing a large model technology, and a charging and discharging scheduling strategy is optimized, so that the adaptability and efficiency of a household storage system are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of energy storage technology, and in particular to a multi-modal household energy storage intelligent scheduling method and system based on large model prediction. Background Art

[0002] With the rapid development of new energy technologies, home energy storage systems (household storage) have gradually become an important part of the power system. Household storage systems can reduce users' electricity costs and improve the overall operating efficiency of the power grid through intelligent scheduling of charging and discharging behaviors. However, the existing intelligent scheduling system has the following shortcomings:

[0003] 1. Limitations of static weights: Current intelligent scheduling strategies usually rely on fixed modal weights, such as electricity prices, power generation, and user electricity usage habits. These weights are usually set based on empirical values ​​and are difficult to reflect dynamic changes in actual scenarios. For example, weather changes will affect photovoltaic power generation, as well as social power consumption and electricity price fluctuations, but traditional methods cannot adjust these weights in real time.

[0004] 2. Lack of real-time optimization capabilities: Existing dispatch systems usually take a long time to obtain dispatch strategies, and are difficult to adapt to rapidly changing environments. In actual applications, dispatch strategies need to be dynamically adjusted based on real-time data to achieve optimal charging and discharging decisions.

[0005] With the rapid development of large model technology, its advantages in processing complex nonlinear relationships, dynamic prediction and real-time optimization have been widely verified. If large model technology is applied to household storage intelligent dispatching system, it can significantly improve the prediction accuracy and dispatching efficiency of the system, thus breaking through the limitations of traditional methods. Summary of the invention

[0006] In view of the above problems and technical requirements, the inventors have proposed a multi-modal household storage intelligent scheduling method and system based on large model prediction. The technical solution of the present invention is as follows:

[0007] In a first aspect, the present application provides a multi-modal household storage intelligent scheduling method based on large model prediction, comprising the following steps:

[0008] Real-time acquisition and preprocessing of multimodal data, including electricity prices, load power demand, photovoltaic power generation and weather data;

[0009] The pre-processed multimodal data is input into the large model to generate the forecast results for each preset period in the future, including the forecast results of electricity price, load power demand and photovoltaic power generation;

[0010] Based on the prediction results of the large model, a multi-objective optimization algorithm is used to generate the initial charging and discharging scheduling strategy for each preset time period;

[0011] The initial charging and discharging scheduling strategy is applied to the household storage system, and the large model is continuously optimized based on the execution effect to roll out the charging and discharging scheduling strategy for future preset time periods.

[0012] A further technical solution is that the large model is implemented using a TFT model, which generates prediction results for each preset period in a future period, including:

[0013] The TFT model divides the preprocessed multimodal data into static features, dynamic historical features, and known future features, and performs embedding operations; static features are data features that remain unchanged in the time series, dynamic historical features are historical data features that change over time, and known future features are known data features within the forecast period;

[0014] The static features, dynamic historical features and known future features after the embedding operation are input into the TFT model. The encoder is used to extract the long-term and short-term dependencies among the three types of features. The decoder is used to combine the dynamic historical features and the known future features to generate the predicted representation, and finally the predicted results are mapped to the target value.

[0015] A further technical solution is that the embedding operation includes:

[0016] Store different features into a vector of fixed length;

[0017] Use the embedding layer to map each vector into a high-dimensional vector, represented as:

[0018] x static =f embed (x static )

[0019] z hist =f embed (x hist )+z time

[0020] z future =f embed (x future )+z time

[0021] Among them, [x static ]、[z static ] represent the original vector and high-dimensional vector storing static features respectively; [x hist ]、[z hist ] represent the original vector and high-dimensional vector storing dynamic historical features respectively; [x future ]、[z future ] represent the original vector and high-dimensional vector storing known future features respectively; fembed () represents the embedding layer function; It represents the periodic code of each preset time period t, and T represents the length of a future period.

[0022] Its further technical solution is to use a multi-objective optimization algorithm to generate the initial charging and discharging scheduling strategy for each preset time period based on the prediction results of the large model, including:

[0023] Get the prediction results for a preset period generated by the large model, including the predicted value of the purchase price C buy (t), predicted value of electricity selling price C sell (t), photovoltaic power generation prediction value P pv (t), load power demand forecast value P load (t);

[0024] Taking the total electricity cost and photovoltaic utilization rate as optimization targets, a multi-objective optimization function f(x) is constructed; the total electricity cost is the difference between the electricity purchase expenditure and the electricity sales income in the preset period, and the photovoltaic utilization rate includes the equivalent value of photovoltaic power supply to loads, photovoltaic power charging to energy storage batteries and photovoltaic power grid feeding in the preset period;

[0025] Establish constraints between the variables to be solved and solve the multi-objective optimization function f(x). The solution form is expressed as:

[0026] Minimize:f(x)

[0027] Subject to:g i (x)≤0,h j (x) = 0

[0028] Among them, x represents the variable to be solved in the function, g i (x) represents the inequality in the constraint condition, h j (x) represents the equality in the constraint condition;

[0029] The optimal solution that satisfies all constraints is obtained as the initial charging and discharging scheduling strategy for the preset period.

[0030] Its further technical solution is that the multi-objective optimization function is:

[0031]

[0032] Among them, the variable to be solved P buy (t), P sell (t), P pv,load (t), P pv,charge (t), P pv,sell(t) are the power purchased by the power grid, the power sold by the power grid, the power supplied by photovoltaic to the load, the charging power of photovoltaic to the energy storage battery, and the photovoltaic power fed into the grid; λ is the weight parameter used to balance the total electricity cost and the photovoltaic utilization rate target; η is the power conversion efficiency; Δt is the duration of the preset period.

[0033] A further technical solution is to establish constraints between the variables to be solved, including:

[0034] Power balance constraint: P pv (t)+P buy (t)+P discharge (t) = P load (t)+P charge (t)+P sell (t);

[0035] Photovoltaic power generation allocation constraint: P pv,load (t)+P pv,charge (t)+P pv,sell (t)≤P pv (t);

[0036] Battery SOC constraints: SOC min ≤SOC(t)≤SOC max ;

[0037] Battery charge and discharge power constraint: 0≤P charge (t)≤P charge,max ,0≤P discharge (t)≤P discharge,max ;

[0038] Dynamic update constraints of battery power:

[0039] Power constraints for buying and selling electricity on the power grid: 0≤P buy (t)≤P buy,max ,0≤P sell (t)≤P sell,max , and all power variables are non-negative;

[0040] Among them, the variable to be solved P discharge P (t) and SOC (t) are the energy storage battery discharge power and battery remaining capacity respectively; charge (t) is the charging power of the energy storage battery; SOC min , SOC max are the minimum remaining battery power and the maximum remaining battery power respectively; P charge,max , P discharge,max are the maximum charging power and the maximum discharging power of the battery respectively; μ is the charging and discharging efficiency; P buy,max , P sell,maxThey are the maximum power of electricity purchased by the power grid and the maximum power of electricity sold by the power grid respectively.

[0041] A further technical solution is to solve the multi-objective optimization function f(x), including:

[0042] The solver selects an initial solution x0, ensures that it satisfies all constraints, and then constructs the Lagrangian:

[0043]

[0044] Among them, λ i and j It is the Lagrange multiplier. The solver adjusts the variables to be solved through iteration, finally reaches convergence and returns the optimal solution.

[0045] Its further technical solution is to continuously optimize the large model according to the execution effect, so as to adjust the charging and discharging scheduling strategy for the future preset time period in a rolling manner, including:

[0046] Divide the future period into multiple rolling windows;

[0047] In each rolling window, the charging and discharging scheduling strategy for the current period is executed, and the prediction results for each preset period in the next future period are updated and iterated based on the multimodal data acquired in real time, so as to dynamically update the charging and discharging scheduling strategy for the next rolling window.

[0048] In a second aspect, the present application also provides a multi-modal household storage intelligent scheduling system based on large model prediction, including:

[0049] Data acquisition and processing module, used to acquire multimodal data in real time and perform preprocessing operations. Multimodal data includes electricity price, load power demand, photovoltaic power generation and weather data;

[0050] The large model prediction module is used to input the pre-processed multimodal data into the large model to generate prediction results for each preset period in the future, including the prediction results of electricity price, load power demand and photovoltaic power generation;

[0051] The strategy optimization module is used to generate the initial charging and discharging scheduling strategy for each preset time period based on the prediction results of the large model using a multi-objective optimization algorithm;

[0052] The system execution and feedback module is used to apply the initial charging and discharging scheduling strategy to the household storage system, and continuously optimize the large model based on the execution effect to roll out the charging and discharging scheduling strategy for future preset time periods.

[0053] The beneficial technical effects of the present invention are:

[0054] In the above method and system, by introducing large model technology to make real-time predictions on multi-modal input variables, and inputting them into the multi-objective optimization function for weight adjustment, an optimized charging and discharging scheduling strategy is generated, which can significantly improve the adaptability and efficiency of the household storage system. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 This is a flow chart of a multi-modal household storage intelligent scheduling method based on large model prediction provided by this application.

[0056] Figure 2 This is a schematic diagram of the composition of a multimodal household storage intelligent scheduling system based on large model prediction provided by this application. DETAILED DESCRIPTION

[0057] The specific implementation of the present invention will be further described below in conjunction with the accompanying drawings.

[0058] In one embodiment of the present application, a multi-modal household storage intelligent scheduling method based on large model prediction is provided. Please refer to Figure 1 As shown, the method comprises the following steps:

[0059] Step 1: Acquire multimodal data in real time and perform preprocessing operations such as cleaning and normalization. The multimodal data includes buying and selling electricity prices, load power demand, photovoltaic power generation and weather data (temperature, light intensity, wind speed, etc.).

[0060] Step 2: Input the preprocessed multimodal data into the large model to generate prediction results for each preset time period in the future, including the prediction results of electricity price, load power demand and photovoltaic power generation.

[0061] Step 3: Based on the prediction results of the large model, a multi-objective optimization algorithm is used to generate the initial charging and discharging scheduling strategy for each preset time period.

[0062] Step 4: Apply the initial charging and discharging scheduling strategy to the household storage system, and continuously optimize the large model based on the execution effect to roll out the charging and discharging scheduling strategy for future preset time periods, thereby further improving the prediction and optimization accuracy.

[0063] In this embodiment, the large model is implemented using the TFT (Temporal Fusion Transformer) model. Since the parameters such as electricity price, load power demand, and photovoltaic power generation have time series characteristics, they need to be combined with time series prediction and external data modeling, which is very suitable for prediction using the TFT large model. The specific implementation method of step 2 includes the following contents: first, the pre-processed multimodal data is divided into three categories:

[0064] 1) Static characteristics (x static): These features are data features that remain unchanged in the time series, such as user type, photovoltaic, energy storage equipment parameters, etc. The feature embedding operation is performed on the feature, specifically including: storing the static features into a fixed-length vector, such as [x static ]; Use the embedding layer function f embed () The vector [x static ] is mapped to a high-dimensional space. Because the original features are sparse and cannot capture the relationship between different parameter values, each category needs to be mapped to a high-dimensional vector through an embedding operation, which is similar to a transformation. The high-dimensional vector of static features [z static ], expressed as:

[0065] z static =f embed (x static )

[0066] 2) Dynamic historical features (x hist ): These features are the historical values ​​of the time series, such as past electricity prices, load power demand, etc. Similarly, these historical data features that change over time are stored in a fixed-length vector, such as [x hist ]; perform the same embedding operation to obtain:

[0067] z hist =f embed (x hist )+z time

[0068] in Indicates the periodic code of each preset period t, where T represents the length of a future period. If t is divided into one hour, then T=24, and if t is divided into 15 minutes, then T=60 / 15×24=96.

[0069] 3) Known future features (x futrue ): These features are known data features within the forecast period, such as timestamps, weather forecast data, etc. These features are also stored in a fixed-length vector, such as [x futrue ]; perform the same embedding operation to obtain:

[0070] z future =f embed (x future )+z time

[0071] Then, the static features, dynamic historical features, and known future features after the embedding operation are input into the TFT model. The encoder is used to extract the long-term and short-term dependencies among the three types of features, and the decoder is used to combine the dynamic historical features and the known future features to generate a predicted representation, and finally the predicted results are mapped to the target value. The entire process can be implemented through the existing TFT-specific prediction algorithm, and only three types of feature data need to be prepared for actual use.

[0072] The entire processing process first defines each preset time period. For example, if each preset time period is 15 minutes, the entire time step is determined. Then the model can output the prediction results of 96 time periods divided by 15 minutes in the next day, and then update the prediction results of the next 96 time periods according to the latest 15 minutes of multimodal data transmitted in real time, so as to ensure that the strategy optimization results in step 3 have the results of medium- and long-term windows and short-term windows sliding. The length of the medium- and long-term window can be understood as the length of the next day or half a day, and the length of the short-term window can be understood as the length of the preset time period, that is, 15 minutes.

[0073] The specific implementation method of step 3 includes the following contents. First, the output of scheduling decision needs to be carried out around the following core goals:

[0074] 1. Economical: Reduce electricity costs by optimizing the timing and power of charging and discharging (such as charging at low electricity prices and discharging at high electricity prices).

[0075] 2. Ensure load power demand: Ensure that load power demand is met first to avoid load power outages or power shortages.

[0076] 3. Maximize the use of renewable energy: Make full use of renewable energy such as photovoltaic power generation to reduce dependence on the power grid.

[0077] According to the prediction results of the large model output in step 2, the prediction results of electricity price, photovoltaic power generation and load power demand in a preset period of time will be provided, for example: the predicted value of purchase price C buy (t): 0.3 yuan / kWh, predicted selling price C sell (t): 0.2 yuan / kWh, photovoltaic power generation forecast value P pv (t): 5.0kW, load power demand forecast value P load (t):3.0kW.

[0078] Then, taking the total electricity cost and photovoltaic utilization rate as the optimization objectives, the multi-objective optimization function f(x) is constructed as follows:

[0079]

[0080] Among them, the first part before λ represents the economic goal, that is, the total electricity cost. The second part after λ represents the photovoltaic utilization rate target, which reduces the total cost by increasing the utilization rate of photovoltaic power generation. λ is a weight parameter used to balance the economic and photovoltaic utilization goals. A larger weight is more inclined to improve the photovoltaic utilization rate, while a smaller weight focuses on economic efficiency. It should be noted that the entire objective function does not simply represent the actual cost. The reason why there are weights is that there may be different tendencies when making the entire goal smaller, such as reducing the electricity cost as much as possible, or maximizing the effectiveness of photovoltaics.

[0081] Among them, P buy (t)·C buy (t)·Δt represents the electricity purchase expenditure in the preset period.

[0082] P sell (t)·C sell (t)·Δt represents the revenue from electricity sales during the preset period. It should be noted that P sell (t) actually includes P pv,sell (t), the only difference is whether the latter will play a greater role after increasing, which needs to be adjusted by the weight parameter λ.

[0083] P pv,load (t)·C buy (t)·Δt represents the equivalent value of the power supplied by photovoltaic power to the load during the preset period.

[0084] P pv,charge (t)·C buy (t)·η·Δt represents the equivalent value of photovoltaic charging the energy storage battery during the preset period. Among them, η is the power conversion efficiency. For example, the amount of electricity allocated by photovoltaics to charge the battery is 1 kWh, but the actual amount of electricity reaching the battery is only 0.9 kWh, and the amount of electricity from the battery to the load is only 0.8 kWh, that is, each layer of conversion has conversion efficiency. This part is because photovoltaics charge the energy storage battery, and the energy storage battery must be used on the load or its electricity is sold in order to realize its value. If there is no photovoltaics, then the battery charging must be through the power grid, so this part of the value lies in the purchase of electricity.

[0085] P pv,sell (t)·C sell (t)·Δt represents the equivalent value of photovoltaic grid feeding within the preset period, where photovoltaic grid feeding refers to the power generation method of directly connecting the electric energy generated by the photovoltaic power generation system to the public power grid after converting it into alternating current through an inverter.

[0086] Secondly, establish the constraints between the variables to be solved, including:

[0087] 1) Power balance constraint: P pv(t)+P buy (t)+P discharge (t) = P load (t)+P charge (t)+P sell (t), this constraint ensures the balance of power supply and demand in each preset period. The left side of the equal sign is the source of power, which is the predicted value of photovoltaic power generation P pv (t), power purchased by the power grid P buy (t), energy storage battery discharge power P discharge (t), the right side of the equal sign is the destination of electricity, which is the load power demand forecast value P load (t), energy storage battery charging power, power grid power sales power P sell (t).

[0088] 2) Photovoltaic power generation allocation constraints: P pv,load (t)+P pv,charge (t)+P pv,sell (t)≤P pv (t), this constraint ensures that the allocation of PV power generation does not exceed the actual power generation.

[0089] 3) Battery status constraints:

[0090] a. Battery SOC constraints: SOC min ≤SOC(t)≤SOC max ,This constraint ensures that the battery power is within the specified minimum remaining battery power and maximum remaining battery power range within a preset period of time.

[0091] b. Battery charge and discharge power constraint: 0≤P charge (t)≤P charge,max ,0≤P discharge (t)≤P discharge,max ,This constraint ensures that within a certain preset period of time, the battery charging and discharging power is always within the specified battery maximum charging power and battery maximum discharging power range.

[0092] c. Dynamic update constraints of battery power: This constraint ensures the dynamic update of the battery charge over time and takes the charge and discharge efficiency μ into account.

[0093] 4) Power constraints for power grid buying and selling: 0≤P buy (t)≤P buy,max ,0≤P sell (t)≤P sell,max This constraint ensures that the power of electricity purchased and sold by the power grid is always within the specified maximum power of the power grid purchased and the maximum power of the power grid sold, and all power variables are non-negative.

[0094] Finally, the multi-objective optimization function f(x) is solved to obtain the optimal solution that satisfies all constraints, which is the initial charging and discharging scheduling strategy for the preset period. The entire solution can be solved by converting it into a mathematical form, which can be expressed in the following form:

[0095] Minimize:f(x)

[0096] Subject to:g i (x)≤0,h j (x) = 0

[0097] Wherein, x represents the variable to be solved in the function, which is expressed as [P buy ,P sell ,P discharge ,SOC,P pv,load ,P pv,charge ,P pv,sell ],g i (x) represents the inequality in the above constraints, h j (x) represents the equality in the above constraint condition.

[0098] This embodiment provides a method for solving a multi-objective optimization function f(x), that is, the method for solving the optimization problem with inequality constraints can use the interior point method, including: the solver selects an initial solution x0 to ensure that it satisfies all constraints, and then constructs the Lagrangian function:

[0099]

[0100] Among them, λ i and j is a Lagrange multiplier. The entire solver will adjust the variable x to be solved through iteration, and finally reach convergence and return the optimal solution. It should be noted that other solution methods can also be used. This application only provides a solution description and does not make specific restrictions here.

[0101] The output conditions of the big model have a relatively large impact on the entire optimization strategy. The key lies in the preset time period Δt. The strategy is optimized for this time period according to the conditions given by the big model. However, the strategy is only for this time period. If you want to achieve global optimization, you need to first give the prediction results of a day divided into different time periods by the big model, provide them to the strategy optimization part, and then correct the strategy based on the real-time prediction data of the new time period. This embodiment adopts a rolling optimization method, that is, first prepare the prediction data and real-time data through steps 1 and 2, where the prediction data is the prediction data for a period of time in the future provided by the big model, including: the purchase price prediction value C of each time period buy (t), predicted value of electricity selling price Csell (t), photovoltaic power generation prediction value P pv (t), load power demand forecast value P load (t), the prediction time range can be one day (e.g. 24 hours), with a time step of 15 minutes. Real-time data includes real-time monitoring of actual data of the current time period, including: current buying and selling electricity prices, photovoltaic power generation, load power demand, battery status, etc., which are used to correct the prediction data.

[0102] Then, the initial global optimization is obtained through step 3, that is, based on the predicted all-day data, a global optimization problem is constructed to obtain a preliminary all-day charging and discharging scheduling strategy, where the objective function used in the global optimization problem is the multi-objective optimization function constructed in step 3. Finally, rolling execution and dynamic optimization are realized through step 4, and its specific implementation method includes:

[0103] Divide a future period into multiple rolling windows. In each rolling window, execute the charge and discharge scheduling strategy of the current period, and update the prediction results of each preset period in the next future period based on the multimodal data acquired in real time, so as to dynamically update the charge and discharge scheduling strategy of the next rolling window. The rolling window of this embodiment can be divided according to the prediction data of the large model, and the rolling optimization is increased to short-term windows and medium- and long-term windows, that is, the full-day data is divided into multiple medium- and long-term windows (for example, half a day or 6 hours), and a medium- and long-term window is divided into multiple short-term windows (for example, 15 minutes). In this way, the calculation can be graded and implemented, thereby realizing real-time changes in the scheduling strategy.

[0104] Based on the same inventive concept, another embodiment of the present application provides a multi-modal household storage intelligent scheduling system based on large model prediction, combined with Figure 2 As shown, it mainly includes the following modules:

[0105] 1) Data acquisition and processing module, used to acquire multimodal data in real time and perform preprocessing operations. Multimodal data includes but is not limited to: buying and selling electricity price data, load power demand data, photovoltaic power generation and weather data (temperature, light intensity, wind speed, etc.).

[0106] 2) Large model prediction module, which is used to input the pre-processed multimodal data into the Transformer-based large model to generate prediction results for each preset period in the future. The core functions of this module include: a. capturing the complex dynamic relationship between multimodal input variables, b. real-time prediction of electricity prices, load power demand and photovoltaic power generation at future times, c. dynamically adjusting the weights of each mode to generate the optimal model input parameters.

[0107] 3) Strategy optimization module, which is used to generate the initial charging and discharging scheduling strategy for each preset time period based on the prediction results of the large model using a multi-objective optimization algorithm. The optimization objectives include: maximizing user benefits (reducing electricity costs and increasing photovoltaic utilization). Optionally, the optimization objectives also include minimizing the burden on the power grid (smoothing the load curve and reducing the peak-to-valley difference).

[0108] 4) System execution and feedback module, which is used to apply the initial charging and discharging scheduling strategy to the household storage system and monitor its execution effect in real time. Through the feedback mechanism, the execution data is returned to the big model for continuous optimization of the model, so as to adjust the charging and discharging scheduling strategy for the future preset period in a rolling manner, further improving the prediction and optimization accuracy.

[0109] Since the implementation solution provided by the system to solve the problem is similar to the implementation solution recorded in the above method, the specific limitations of each module can be found in the limitations of the corresponding steps in the multi-modal household storage intelligent scheduling method above, and will not be repeated here.

[0110] The above is only a preferred embodiment of the present application, and the present invention is not limited to the above embodiments. It is understood that other improvements and changes directly derived or associated by those skilled in the art without departing from the spirit and concept of the present invention should be considered to be included in the protection scope of the present invention.

Claims

1. A multi-modal household storage intelligent scheduling method based on large model prediction, characterized in that: The method comprises: Acquire multimodal data in real time and perform preprocessing operations, wherein the multimodal data includes electricity price, load power demand, photovoltaic power generation and weather data; The pre-processed multimodal data is input into the large model to generate the forecast results for each preset period in the future, including the forecast results of electricity price, load power demand and photovoltaic power generation; Based on the prediction results of the large model, a multi-objective optimization algorithm is used to generate the initial charging and discharging scheduling strategy for each preset time period; The initial charging and discharging scheduling strategy is applied to the household storage system, and the large model is continuously optimized according to the execution effect to roll-over the charging and discharging scheduling strategy for the future preset time period.

2. The multi-modal household storage intelligent scheduling method based on large model prediction according to claim 1 is characterized in that: The large model is implemented using the TFT model, which generates prediction results for each preset time period in a future period, including: The TFT model divides the preprocessed multimodal data into static features, dynamic historical features and known future features, and performs an embedding operation; wherein the static features are data features that remain unchanged in the time series, the dynamic historical features are historical data features that change over time, and the known future features are known data features within the prediction time period; The static features, dynamic historical features and known future features after the embedding operation are input into the TFT model, the encoder is used to extract the long-term and short-term dependencies among the three types of features, and the decoder is used to combine the dynamic historical features and the known future features to generate a predicted representation, and finally the predicted results are mapped to the target value.

3. The multi-modal household storage intelligent scheduling method based on large model prediction according to claim 2 is characterized in that: The embedding operation includes: Store different features into a vector of fixed length; Use the embedding layer to map each vector into a high-dimensional vector, represented as: z static =f embed (x static ) z hist =f embed (x hist )+z time z future =f embed (x future )+z time Among them, [x static ]、[z static ] represent the original vector and high-dimensional vector storing static features respectively; [x hist ]、[z hist ] represent the original vector and high-dimensional vector storing dynamic historical features respectively; [x future ]、[z future ] represent the original vector and high-dimensional vector storing known future features respectively; f embed () represents the embedding layer function; It represents the periodic code of each preset time period t, and T represents the length of a future period.

4. The multi-modal household storage intelligent scheduling method based on large model prediction according to claim 1 is characterized in that: The prediction results based on the large model are used to generate the initial charge and discharge scheduling strategy for each preset time period using a multi-objective optimization algorithm, including: Get the prediction results for a preset period generated by the large model, including the predicted value of the purchase price C buy (t), predicted value of electricity selling price C sell (t), photovoltaic power generation prediction value P pv (t), load power demand forecast value P load (t); The multi-objective optimization function f(x) is constructed with the total electricity cost and photovoltaic utilization rate as optimization targets; wherein the total electricity cost is the difference between the electricity purchase expenditure and the electricity sales income in the preset period, and the photovoltaic utilization rate includes the equivalent value of photovoltaic power supply to loads, photovoltaic power charging to energy storage batteries and photovoltaic power grid feeding in the preset period; Establish constraints between the variables to be solved, and solve the multi-objective optimization function f(x). The solution form is expressed as: Minimize:f(x) Subject to:g i (x)≤0,h j (x)=0 Among them, x represents the variable to be solved in the function, g i (x) represents the inequality in the constraint condition, h j (x) represents the equality in the constraint condition; The optimal solution that satisfies all constraints is obtained as the initial charging and discharging scheduling strategy for the preset period.

5. The multi-modal household storage intelligent scheduling method based on large model prediction according to claim 4 is characterized in that: The multi-objective optimization function is: Among them, the variable to be solved P buy (t), P sell (t), P pv,load (t), P pv,charge (t), P pv,sell (t) are the power purchased by the power grid, the power sold by the power grid, the power supplied by photovoltaic to the load, the charging power of photovoltaic to the energy storage battery, and the photovoltaic power fed into the grid; λ is the weight parameter used to balance the total electricity cost and the photovoltaic utilization rate target; η is the power conversion efficiency; Δt is the duration of the preset period.

6. The multi-modal household storage intelligent scheduling method based on large model prediction according to claim 4 is characterized in that: The step of establishing constraint conditions between variables to be solved includes: Power balance constraint: P pv (t)+P buy (t)+P discharge (t) = P load (t)+P charge (t)+P sell (t); Photovoltaic power generation allocation constraint: P pv,load (t)+P pv,charge (t)+P pv,sell (t)≤P pv (t); Battery SOC constraints: SOC min ≤SOC(t)≤SOC max ; Battery charge and discharge power constraint: 0≤P charge (t)≤P charge,max ,0≤P discharge (t)≤P discharge,max ; Dynamic update constraints of battery power: Power constraints for buying and selling electricity on the power grid: 0≤P buy (t)≤P buy,max ,0≤P sell (t)≤P sell,max , and all power variables are non-negative; Among them, the variable to be solved P discharge P (t) and SOC (t) are the energy storage battery discharge power and battery remaining capacity respectively; charge (t) is the charging power of the energy storage battery; SOC min , SOC max are the minimum remaining battery power and the maximum remaining battery power respectively; P charge,max , P discharge,max are the maximum charging power and the maximum discharging power of the battery respectively; μ is the charging and discharging efficiency; P buy,max , P sell,max They are the maximum power of electricity purchased by the power grid and the maximum power of electricity sold by the power grid respectively.

7. The multi-modal household storage intelligent scheduling method based on large model prediction according to claim 4 is characterized in that: Solving the multi-objective optimization function f(x) includes: The solver selects an initial solution x0, ensures that it satisfies all constraints, and then constructs the Lagrangian: Among them, λ i and j is a Lagrange multiplier. The solver adjusts the variables to be solved through iteration, finally converges and returns the optimal solution.

8. The multi-modal household storage intelligent scheduling method based on large model prediction according to claim 1 is characterized in that: The large model is continuously optimized according to the execution effect to adjust the charging and discharging scheduling strategy for the future preset time period in a rolling manner, including: Divide the future period into multiple rolling windows; In each rolling window, the charging and discharging scheduling strategy for the current period is executed, and the prediction results for each preset period in the next future period are updated and iterated based on the multimodal data acquired in real time, so as to dynamically update the charging and discharging scheduling strategy for the next rolling window.

9. A multi-modal household storage intelligent scheduling system based on large model prediction, characterized in that: include: A data acquisition and processing module, used to acquire multimodal data in real time and perform preprocessing operations, wherein the multimodal data includes electricity price, load power demand, photovoltaic power generation and weather data; The large model prediction module is used to input the pre-processed multimodal data into the large model to generate prediction results for each preset period in the future, including the prediction results of electricity price, load power demand and photovoltaic power generation; The strategy optimization module is used to generate the initial charging and discharging scheduling strategy for each preset time period based on the prediction results of the large model using a multi-objective optimization algorithm; The system execution and feedback module is used to apply the initial charge and discharge scheduling strategy to the household storage system, and continuously optimize the large model according to the execution effect to roll-adjust the charge and discharge scheduling strategy for the future preset time period.

Images