Power distribution network air conditioner load user preference learning and operation strategy generation method

Abstract

The invention discloses a power distribution network air conditioner load user preference learning and operation strategy generation method, relates to the field of power distribution networks, and aims to quickly and actively learn user preferences based on limited and random user behavior observation. On the basis of learned user preferences, an optimal temperature setting plan in several hours in the future is generated by solving a proposed model based on a multi-cycle random expected value; an operation strategy capable of closely tracking changes of electricity price and outdoor temperature forecast and changes of user preferences is designed, the changes of the user preferences can be well and dynamically tracked, and a temperature setting plan is automatically updated; finally, the thermostat can be operated with maximized user benefits.

Description

Technical Field

[0001] The invention relates to the field of distribution networks, and in particular to a method for learning user preferences of air-conditioning loads in distribution networks and generating operating strategies. Background Art

[0002] Smart thermostats can automatically adjust indoor temperature by controlling air conditioners according to user preferences, saving electricity without affecting comfort. However, there is a conflict between providing comfortable temperature and saving electricity when the thermostat is working, so it is crucial to develop optimal operation strategies for thermostats that can cope with different user preferences. The method proposed in the present invention is deployed on air-conditioning smart thermostats, such as Nest, Honeywell, and Ecobee. The present invention learns finite and random user preferences through quantile regression, and based on the learned user preferences, generates the optimal temperature setting plan by solving stochastic expected value models (SEVM). According to the designed operation strategy, it can dynamically track changes in user preferences and automatically update the temperature setting plan. First, quantile regression and stochastic expected value models (SEVM) are introduced respectively:

[0003] 1) Quantile regression

[0004] The ordinary linear regression (OLS) model focuses on the mean and studies the expected mean of the response variable under the condition that some explanatory variables are fixed. The model estimation method is the least squares method, which minimizes the sum of squares of the residuals of each sample. However, the most basic assumption of linear regression is that the residuals satisfy the normal distribution, random independence, and equal variance. These conditions are often not met in reality. If there are outliers in the sample data, the estimated value of the linear regression model may have a large deviation. Sometimes we not only want to study the expected mean of the response variable, but also want to explore the global distribution of the response variable (such as a certain quantile of the response variable). At this time, quantile regression is needed. The application conditions of quantile regression are relatively loose, which can describe the global characteristics of the response variable and can mine more abundant information. In addition, quantile regression estimation uses the weighted least absolute deviation (WLAD) method for estimation, which is usually not affected by outliers and the results are more robust. Corresponding to the least squares method of the residual sum of squares, the minimum absolute deviation sum is also called the least squares method.

[0005] After a set of data is arranged from small to large, if the q quantile is m, it means that 100q% of the data in the set is less than m. Quantile regression combines linear regression with the concept of quantile. Putting aside the difficult-to-understand mathematical formulas, in layman's terms, the so-called q quantile regression means that we hope that there are 100q% data points under the fitting line, such as 25% of the data points under the 0.25 quantile regression line. Therefore, series quantile regression does not fit a curve like linear regression, but can fit a cluster of curves. Different regression coefficients for different quantiles indicate that the explanatory variables have different effects on the response variables at different levels. We can use this to obtain the influence of the explanatory variables on the changing trend of the quantiles of the response variables.

[0006] 2) Random Expected Value Model

[0007] As a subclass of stochastic programming, stochastic expected value models (SEVM) are a type of stochastic programming that aims to optimize the expected value of the objective function under the constraints of stochastic expected values, thereby providing the decision maker with the best choice.

[0008] The existing technical solutions proposed a framework for optimal temperature set point decision of smart thermostat based on support vector machines (SVM). This solution has low computational complexity, making it very suitable for implementation as an embedded system. This section introduces the proposed framework to support smart thermostat decision making, such as Figure 1 The input raw data comes from: 1) weather sensors (indoor / outdoor temperature and humidity values, solar radiation); 2) occupant activities (such as the presence and activity in the study room); 3) building sensors (data from energy metering devices).

[0009] The results of the sensors are temporarily stored to support the training task of the support vector machine. In this step, two complementary mechanisms, called coarse-grained windowing and fine-grained windowing, are used, respectively, in order to minimize the amount of data without affecting the quality of the derived training. More specifically, instead of saving the raw data acquired from the sensors, the coarse-grained window selectively stores the data corresponding to the last x days. A further improvement is achieved by emphasizing (using fine-grained windows) only up to a certain duration per day. Thus, instead of operating on the entire historical data, only a small part of it is required (see Figure 2 ).

[0010] Both the number of preceding days and the duration of each day are dynamically customizable to better utilize hardware resources. As we will discuss later, this technique, in addition to accelerating enhancements, significantly reduces the demand for onboard processing power and storage requirements. In contrast to related studies of decision-making mechanisms that rely on statistical analysis, the proposed framework considers only a small subset of historical data and weather forecasts during its training phase, resulting in near-optimal results. Furthermore, the approach does not apply any iterative and / or time-consuming simulations to calculate the temperature set points. In other words, the almost negligible computational complexity imposed by the decision-making framework facilitates its implementation as part of low-cost embedded systems (e.g., smart thermostats).

[0011] The temperature set points of each thermal zone are calculated using a support vector machine mechanism; a supervised learning model with an associated learning algorithm is used to analyze the data for regression analysis. Similar to the associated artificial intelligence algorithms, the training of SVMs introduces considerable computational and storage complexity, which is not always feasible for embedded devices. To overcome this limitation, appropriate techniques can be used, such as the introduction of sliding windows, e.g. Figure 2 shown.

[0012] In order to be able to continue retraining and thus improve the efficiency of the system. The SVM used is retrained once for each calculated temperature set point. The output of the updated (trained) SVM are two models for estimating the energy consumption of each thermal zone as well as the predicted thermal comfort. While energy costs can be easily quantified with metering devices, the thermal comfort of the occupants can only be estimated. Various approaches can be used for this purpose, and the thermal comfort model developed by Fanger was adopted. More precisely, the model relates environmental and physiological factors to thermal sensations to estimate the predicted percentage of dissatisfied people in the room. The problem of determining the optimal temperature set point for each thermal zone cannot be considered trivial due to the intermittent behavior of solar energy, the uncertain dynamics of buildings and the need to meet the thermal comfort constraints of the microgrid occupants. This is a typical multi-objective problem, where the reduction of energy costs and the maximization of thermal comfort are in conflict with each other. Therefore, there is no single optimal solution. Instead, a set of trade-offs must be calculated in order to obtain the best possible solution between the objectives.

[0013] The disadvantages of the prior art are:

[0014] 1) Lack of personalized and dynamic evaluation of user preferences

[0015] User preference refers to the trade-off between electricity costs and thermal comfort that users make. In many studies, it is simply assumed that user temperature settings vary within a certain "comfort range" or simulated using a unified model rather than a user-specific model. But in practice, preferences between different users may vary greatly and may change over time. Therefore, it is also crucial to dynamically estimate user preferences while taking into account the randomness of user behavior.

[0016] 2) Unrealistic requirements on the user side

[0017] There are some AI approaches to learning user preferences, but these studies often require users to adjust their thermostats frequently to get enough observations to train the AI ​​model, which requires users to have some programming skills. Summary of the invention

[0018] The purpose of the present invention is to provide a method for learning user preferences and generating operation strategies for air-conditioning loads in a distribution network, so as to solve the problems raised in the above-mentioned background technology.

[0019] To achieve the above object, the present invention provides the following technical solutions:

[0020] A method for learning user preferences and generating operation strategies for air-conditioning loads in a distribution network comprises the following steps:

[0021] Step 1) Estimate user preferences based on quantile regression method:

[0022] By estimating the temperature set point T set The probability relationship between the user's preference and the electricity price e is used to estimate the user's preference; first, the user's total cost function C all Defined as the purchase cost C pur (C pur =eq) and discomfort level C dis The sum of T ideal The discomfort function C is constructed using the sixth-order Taylor polynomial as the center dis w=(w1,w2,...,w6) is the discomfort coefficient;

[0023] Then the temperature setting T is expressed by the thermodynamic equation of the building. set Relationship with power consumption q:

[0024]

[0025] Where m is the order of the Taylor polynomial, w m is the mth component of the discomfort coefficient w, λ is the building heat dissipation coefficient, β is the thermal parameter of the building, Δt is the time interval, T0 is the initial indoor temperature, q is the power consumption, Tout is the outdoor temperature; the optimal solution of the above formula should satisfy the Karush-Kuhn-Tucher (KKT) condition, and further obtain:

[0026]

[0027] Then, based on the above formula (2), the user's temperature setting set and electricity price e, we get the joint probability density distribution of w = (w1, w2, ..., w6);

[0028] When more evidence or information is available, quantile regression is applied to update the probability of the hypothesis, and the probability distribution of the discomfort coefficient w is obtained by quantile regression to estimate the temperature set point T set The probability relationship between the electricity price e is as follows:

[0029] Define a prior distribution to express the "degree of belief" in the possible values ​​of w:

[0030]

[0031] In the formula, α is a hyperparameter, the prior and likelihood are both Gaussian distributions, so the posterior is also Gaussian distribution:

[0032] p(w|e ob ,α,σ 2 )=N(μ w ,∑) (4)

[0033] In the formula, e ob is the electricity price vector, σ 2 is the Gaussian noise ε n The variance of w The mean of the posterior vector is the ∑ posterior covariance matrix, defining the matrix Φ nm =(T n -T ideal ) m Then we have:

[0034]

[0035] Choose α=α MP and in is the marginal likelihood p(e ob |α,σ 2 ):

[0036]

[0037] Where I is the unit matrix and N is the number of observations. This maximization problem is solved by particle swarm optimization. The optimal parameters are found through Occam’s razor, and then the joint probability density distribution of w is obtained. In summary, we can get:

[0038]

[0039] Where e est It is T set.est Estimation of the electricity price, μ e is the mean of the marginal distribution, is the variance of the marginal distribution, and the other parameters are the same as above; according to formula (7), we estimate T set The probability relationship between it and the electricity price e;

[0040] Step 2), generate the optimal temperature setting plan based on the random expectation model:

[0041] When the discomfort coefficient w is determined, the optimization problem in equation (1) is solved to generate the optimal temperature setting schedule. In order to obtain a more practical and comprehensive schedule, a single cycle is used to further discuss equation (1):

[0042]

[0043] Where T sche is the temperature setting table, T min is the lower limit of the temperature setting range, T max is the upper limit of the allowable range of temperature setting, and the other parameters are the same as above. In order to further reduce the electricity bill by utilizing the heat storage capacity of the building under the real-time electricity price, the multi-period random expectation model of formula (1) is established in formula (9):

[0044]

[0045] Where e t is the electricity price forecast at time t, q t is the power consumption at time t, q max is the rated capacity of the high voltage AC system, T sche.t is the temperature setting schedule at time t, T out.t is the outdoor temperature forecast at time t, T max It is the upper limit of the allowable range of temperature setting, and the other parameters are the same as above;

[0046] Then, the Monte Carlo simulation is used to solve equation (9). In time period t, equation (9) is solved to obtain the temperature setting plans for time period t+1, time period t+2, ..., time period t+T, which will minimize the expected total cost of T time periods. Since the multi-period random expectation model can produce a more economical temperature setting plan than the single-period random expectation model, the thermostat operation strategy proposed in the subsequent step 3 will be based on equation (9);

[0047] Step 3), a method for generating a thermostat automatic operation strategy based on user preference estimation and temperature setting scheduling:

[0048] Based on the research of step 1 user preference estimation and step 2 temperature setting scheduling, a method for generating automatic operation strategy for thermostat is designed, and the operation strategy is divided into three stages: startup stage, user preference estimation stage and automatic operation stage; in the startup stage, users specify their ideal temperature and select the default mode to start the thermostat; in the user preference estimation stage, the thermostat will estimate user preferences based on user behavior observations and generate a temperature setting schedule in a timely manner; in the automatic operation stage, the outdoor temperature and real-time electricity price in the next T hours are first predicted, and the random expectation model is solved by Monte Carlo simulation to generate a temperature setting plan for the next T hours, and the thermostat will execute the scheduling in the next 1 hour while continuing to detect user input. If no new user input is detected in the next 1 hour, the thermostat will repeat the process of this stage.

[0049] Compared with the prior art, the beneficial effects of the present invention are: the present invention can quickly and actively learn user preferences based on limited and random user behavior observations; and based on the learned user preferences, generate the optimal temperature setting plan for the next few hours by solving a proposed multi-period random expected value model; and design an operation strategy that can closely track changes in electricity prices and outdoor temperature forecasts as well as changes in user preferences, which can dynamically track changes in user preferences and automatically update the temperature setting plan; ultimately, the thermostat can be operated to maximize user benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is a framework diagram for calculating the temperature set point in the prior art.

[0051] Figure 2 A conceptual diagram of coarse-grained and fine-grained sliding windows in the prior art.

[0052] Figure 3 It is the potential preference of the virtual user in the present invention.

[0053] Figure 4 It is a comparison diagram of marginal likelihood and test error under different combinations in the present invention.

[0054] Figure 5 These are the fitting results of least squares fitting and quantile regression under different α values ​​in the present invention.

[0055] Figure 6 This is a probability estimation diagram of user preferences under three different observation values ​​in the present invention.

[0056] Figure 7Schematic diagram of the expected discomfort function in the present invention.

[0057] Figure 8 This is a schematic diagram of the optimal temperature schedule and corresponding power consumption for three users in the present invention.

[0058] Fig. 9 It is a schematic diagram for comparing the total costs of the fixed set value plan and the recommended plan in the present invention.

[0059] Fig.10 1 is an update diagram of the temperature setting schedule under situation (1) in an embodiment of the present invention.

[0060] Fig.11 4 is an update diagram of the temperature setting schedule under situation (2) in the embodiment of the present invention.

[0061] Fig.12 Schematic diagram of re-estimating user preferences based on new observation results in an embodiment of the present invention

[0062] Fig.13 This is a flow chart of the operation strategy of the present invention.

[0063] Fig.14 This is a pseudo code diagram of the operation strategy of the present invention. DETAILED DESCRIPTION

[0064] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0065] In an embodiment of the present invention, a method for learning user preferences and generating operation strategies for air-conditioning loads in a distribution network includes the following steps:

[0066] Step 1), estimate user preferences based on the quantile regression method.

[0067] As a preferred embodiment, the method of step 1) is:

[0068] By estimating the temperature set point T set The probability relationship between the user's price and the electricity price e (that is, the user's preference) is used to estimate the user's preference. First, the user's total cost function C all Defined as the purchase cost C pur (C pur =eq) and discomfort level C dis and we use the sixth-order Taylor polynomial (with T ideal The discomfort function is constructed by w=(w1,w2,...,w6) is the discomfort coefficient.

[0069] Then the temperature setting T is expressed by the thermodynamic equation of the building. set Relationship with power consumption q:

[0070]

[0071] Where m is the order of the Taylor polynomial, w m is the mth component of the discomfort coefficient w, λ is the building heat dissipation coefficient, β is the thermal parameter of the building, Δt is the time interval, T0 is the initial indoor temperature, q is the power consumption, T out is the outdoor temperature. The optimal solution of the above equation should satisfy the Karush-Kuhn-Tucher (KKT) condition, then we can get:

[0072]

[0073] Then, based on the above formula (2), the user's temperature setting set and electricity price e, the proposed thermostat can obtain a joint probability density distribution of w = (w1, w2, ..., w6).

[0074] When more evidence or information is available, quantile regression can be applied to update the probability of the hypothesis. Here we use quantile regression to obtain the probability distribution of the discomfort coefficient w and estimate the temperature set point T set The probabilistic relationship between the electricity price e (that is, the user preference) is as follows:

[0075] We define a prior distribution to express our “belief” in the possible values ​​of w:

[0076]

[0077] In the formula, α is a hyperparameter, the prior and likelihood are both Gaussian distributions, so the posterior is also Gaussian distribution:

[0078] p(w|e ob ,α,σ 2 )=N(μ w ,∑) (4)

[0079] In the formula, e ob is the electricity price vector, σ 2 is the Gaussian noise ε n The variance of w The mean of the posterior vector is the ∑ posterior covariance matrix. Define the matrix Φ nm =(T n -T ideal )m Then we have:

[0080]

[0081] We choose α = α MP and in is the marginal likelihood p(e ob |α,σ 2 ):

[0082]

[0083] In the formula, I is the unit matrix and N is the number of observations. This maximization problem can be solved by particle swarm optimization. The optimal parameters can be effectively found through Occam's razor, and then the joint probability density distribution of w can be obtained. In summary, we can get:

[0084]

[0085] In the formula, e est It is T set.est Estimation of the electricity price, μ e is the mean of the marginal distribution, is the variance of the marginal distribution, and the other parameters are the same as above. According to formula (7), we can estimate T set The probabilistic relationship between and the electricity price e (that is, user preference).

[0086] Step 2), generate the optimal temperature setting plan based on the random expectation model.

[0087] As a preferred embodiment, the method of step 2) is:

[0088] When the discomfort coefficient w is determined, the optimization problem in equation (1) can be solved to generate the optimal temperature setting schedule. However, in order to obtain a more practical and comprehensive schedule, we further discuss equation (1) using a single cycle:

[0089]

[0090] Where T sche is the temperature setting table, T min is the lower limit of the temperature setting range, T max is the upper limit of the temperature setting range, and the other parameters are the same as above. In order to further reduce the electricity bill by utilizing the heat storage capacity of the building under the real-time electricity price, we established the multi-period random expectation model of formula (1) in formula (9):

[0091]

[0092] Where e tis the electricity price forecast at time t, q t is the power consumption at time t, q max is the rated capacity of the high voltage AC system, T sche.t is the temperature setting schedule at time t, T out.t is the outdoor temperature forecast at time t, T max It is the upper limit of the allowable range of temperature setting, and the other parameters are the same as above.

[0093] Then, by solving equation (9) through Monte Carlo simulation, in time period t, solving equation (9) can obtain the temperature setting plans for time period t+1, time period t+2, ..., time period t+T, which will minimize the expected total cost of T time periods. Since the multi-period random expectation model can produce a more economical temperature setting plan than the single-period random expectation model, the thermostat operation strategy we propose next will be based on equation (9).

[0094] Step 3), automatic operation strategy of the thermostat is generated based on user preference estimation and temperature setting scheduling.

[0095] As a preferred embodiment, the method of step 3) is:

[0096] Based on the research of step 1 user preference estimation and step 2 temperature setting scheduling, we can design a method for generating automatic operation strategy for thermostat. This paper mainly studies the operation strategy of the thermostat in automatic mode when the user is at home, and divides the operation strategy into three stages: startup stage, user preference estimation stage and automatic operation stage. In the startup stage, users can specify their ideal temperature and select the default mode to start the thermostat. In the user preference estimation stage, the thermostat will estimate the user preference based on the user behavior observation and generate the temperature setting schedule by itself in a timely manner. In the automatic operation stage, the outdoor temperature and real-time electricity price in the next T hours are first predicted, and the random expectation model is solved by Monte Carlo simulation to generate the temperature setting plan for the next T hours. And the thermostat will execute the scheduling in the next 1 hour while continuing to detect user input. If no new user input is detected in the next 1 hour, the thermostat will repeat the process of this stage. In summary, after the startup stage, the thermostat cycles between "learning" and "scheduling". When user input is detected, it updates the user preferences and generates a temperature setting plan based on the updated user preferences. The operation strategy flow chart is as follows Fig.13 The corresponding algorithm pseudo code is shown as Fig.14 shown.

[0097] The simulation analysis is as follows:

[0098] We examine the performance of the user preference estimation method with predefined “virtual users”. Then, the proposed thermostat operation strategy is simulated under the real-time electricity price in Washington to verify its effectiveness.

[0099] Hypothetical user case:

[0100] 1) Generate a "virtual user"

[0101] The behavior of predefined "virtual users" is as follows Figure 3 As shown, the vertical and horizontal axes represent different electricity prices and the corresponding observed user temperature settings, respectively. The red dashed line represents the potential preference of the "virtual user" in summer. Generally speaking, the higher the electricity price, the higher the temperature setting. In reality, almost no one knows this red dashed line except the user himself. In fact, due to the user's subjective feelings about thermal comfort and price sensitivity, even the user himself cannot accurately describe his preferences, nor can he make every decision based on his potential preferences accurately. Fortunately, some observations of her temperature settings under a range of electricity prices are available. For the defined "virtual user", these observations are generated through simulation, such as Figure 3 The goal of this part is to recover the true preferences of the “virtual user” (red dashed line) based on the limited observations (asterisks).

[0102] 2) Optimize hyperparameters α and σ 2

[0103] As mentioned above, the distribution of w can be obtained by equations (4) and (5). However, the hyperparameters α and σ 2 is unknown and must be estimated first. Here, we will find α and σ that maximize the marginal likelihood in (6) 2 The optimization results of the particle swarm algorithm under different observation times are shown in Table 1.

[0104] Table 1 Optimal values ​​of hyperparameters that maximize marginal likelihood

[0105] Tab.1 Optimal values ​​of the hyper-parameters that maximize the marginal likelihood

[0106]

[0107] To examine the performance of the selected hyperparameters, we compared 2 The marginal likelihood (calculated using formula (6)) and test error E under different combinations of test . Use e=f U (T set ) represents the underlying user preference function ( Figure 3 The red dashed line in the middle), the test error is defined as:

[0108]

[0109] That is, given a (α,σ 2 ), the test error reflects the difference between the proposed Taylor polynomial-based user preference function and the underlying user preference f U The degree of closeness. 2 ) ML and (α,σ 2 ) TE denote the (α,σ) that produces the maximum marginal likelihood and the minimum test error respectively. 2 ). (α,σ 2 ) ML can be completely obtained from user behavior observation, and (α,σ 2 ) TE Only given the user's potential preference f U An effective (α,σ 2 ) ML should be consistent with (α,σ 2 ) TE Tightly integrated.

[0110] We take these 10 observations as an example, and the results are as follows Figure 4 As shown. We can see that (α,σ 2 ) ML The test error is very small, 2.0588$ / MWh. 2 ) TE =(0.01,36), the actual minimum test error is 2.0504$ / MWh, which is close to (α,σ 2 ) ML =(0.03,45). The user preference model based on Taylor polynomials proposed under the condition of large α can only fit the smooth underlying user preference function (i.e. Figure 5 The red dashed line in the figure fits the data generated by the smooth and complex underlying user preference function under the condition of small α. However, due to the marginal likelihood p(e ob |α,σ 2 ) is the normalized probability distribution over all possible data sets. The improvement of Taylor polynomial fitting ability will lead to a decrease in the probability of finding potential preferences. The maximum marginal likelihood is actually the selection of the simplest Taylor polynomial model among all models that can explain the observed values, which is an automatic application of Occam's Razor.

[0111] Figure 5The performance of the proposed estimation method is shown. Based on the user preference model in Eq. (7), the least squares fitting (equivalent to α = 0) and α = 10 -4 The models only focus on fitting accuracy. As a result, they explain the observations in a complex way, both of which lead to overfitting. In contrast, when α = 10 2 When α = 0.03, the model is too smooth and cannot explain the observations. When α = 0.03, the marginal likelihood of the model is maximized, which controls the complexity of the model while explaining the observations. Therefore, the model with α = 0.03 is most likely to be a potential preference. In short, quantile regression can select the optimal model complexity by maximizing the marginal likelihood. This can avoid overfitting and ensure the learning ability under a small number of observations, which is the theoretical basis of "fast learning".

[0112] 3) Estimating Probabilistic User Preferences

[0113] After determining the optimal hyperparameters by maximizing the marginal likelihood, we can obtain the probability distribution of w under 10, 30, and 60 observations. Then we transform w, α, and σ 2 Marginalization is used to estimate probabilistic user preferences, such as Figure 6 As shown in Figure 2. We can see that as the number of observations increases from 10 to 30 and 60, the estimation becomes more focused on the underlying user preferences. It is also worth noting that some outliers are excluded from the confidence interval. This feature suggests that this method can ignore the user's unexpected behavior and recover the user's true preferences.

[0114] Real user cases:

[0115] A simulation of real-time pricing was conducted in Washington, DC, to verify the effectiveness of the proposed method in improving user satisfaction. We made the following two simplifications: ignoring the initial stage including the default mode and replacing the predicted values ​​with actual values ​​(the prediction technology of electricity prices and outdoor temperature is not within the scope of this invention). It takes about 3.7 seconds to run the entire model on a personal laptop (i58250U, 8G). Given that the proposed strategy runs once an hour and does not require a real-time solution, these results can support the claim that it is technically feasible to implement the proposed method on an embedded smart thermostat.

[0116] 1) Estimating user preferences

[0117] We randomly select 3 users (represented by users A, B, and C) from the historical temperature settings of 50 users in Washington, DC, and set 10 temperatures respectively to form our observations and the corresponding real-time electricity prices. The observations are listed in Table 2. The expected discomfort function (all possible values ​​of w) is estimated and shown in Figure 7 middle.

[0118] Table 2 Observations of three users in Washington, DC

[0119] Tab.2 Observations of three users in WASHINGTON DC

[0120]

[0121] It can be seen that the three users have different comfort zones, within which their discomfort levels are close to 0. Moreover, the three users have different price sensitivities: User A has a relatively wide comfort zone and does not pay much attention to comfort, which means that she has a high tolerance for temperature changes and is more sensitive to electricity costs; on the contrary, User C has higher requirements for thermal comfort and is less sensitive to prices. Table 3 summarizes the characteristics of user preferences.

[0122] Table 3 Characteristics of user preferences

[0123] Tab.3 Characteristics of user's preferences

[0124]

[0125] 2) Generate temperature setting schedule:

[0126] The optimal temperature setting schedule from 6:00 to 17:00 on August 6, 2017 is generated by solving the random expectation model SEVM in equation (9). Assume that the initial indoor temperature is equal to the outdoor temperature. The thermal parameters and high-voltage AC parameters are shown in Table 4. The optimal temperature scheduling and corresponding power consumption of the three users are shown in Figure 8 As shown. It can be seen that user B gets the schedule with the lowest temperature setting among the three users, and therefore consumes the most electricity because her comfort zone is cooler and she is not so sensitive to electricity prices. User A gets the most energy-efficient plan because she is price-sensitive and has a high tolerance for temperature changes. Although user C has a high requirement for thermal comfort, his energy consumption is moderate because he prefers a relatively warm environment. Figure 8 The schedule in (1) keeps the temperature roughly within the comfort range of the three users to avoid causing considerable discomfort to the users.

[0127] Table 4 Thermal and high voltage AC parameters of the building

[0128] Tab.4 Thermodynamic and HVAC parameters of the building

[0129]

[0130] The peak electricity consumption period is 12:00-13:00. This is because the electricity price is expected to increase significantly in the next few hours. Therefore, the proposed strategy chooses to pre-cool the house during the period of 12:00-13:00 when the price is relatively low. In this way, the house can store coolness to reduce the electricity consumption in the next few peak hours and minimize the total cost during the entire scheduling period. To further demonstrate the effectiveness of the proposed method, we will Figure 8 The total cost of the scheduling plan generated in (1) is compared with the traditional fixed set point mode. The total cost of the three users in different modes is shown in Fig. 9 and as shown in Table 5. The results show that the minimum costs of the three users under the fixed set point scheme are higher than the proposed optimal scheme, and the method proposed in the invention can save at least 12% of the total cost on average.

[0131] Table 5 Total cost comparison between fixed setpoint schedule and proposed schedule

[0132] Tab.5 Overall cost comparison between fixed set-point and the proposed schedule

[0133]

[0134] 3) Rolling Update:

[0135] The above schedule is generated at 6:00 and executes the schedule from 6:00 to 7:00. Figure 3 According to the operation strategy of , the temperature setting plan is updated in the following two cases: 1 hour after the last temperature setting table is generated or new user input is detected. Let's take user A as an example: Assume that the user of the above building sets the thermostat to 23℃ at 11:36. Figure 8 For the temperature setting schedule in (1), we demonstrate the subsequent update of the schedule in two cases:

[0136] Case (1) (6:00-11:00): From 6:00 to 7:00, the thermostat detects no user input. Therefore, at 7:00, the thermostat predicts the electricity price and outdoor temperature from 7:00 to 18:00, and then automatically generates a new temperature setting schedule for this period. It should be noted that only the temperature setting for the time period 7:00 to 8:00 is executed here. Similarly, because user input is not detected until 11:36, from 8:00 to 11:00, the thermostat automatically updates the temperature setting schedule at the beginning of each hour and implements the temperature setting in the first hour of each schedule. The user preference is not updated until 11:36. In case (1), the temperature setting schedule is updated as follows: Fig.10 shown.

[0137] Case (2) (11:00-12:00): At 11:00, the thermostat automatically generates a temperature setting schedule from 11:00 to 22:00 ( Fig.11 ). At 11:36, the user input 23°C is detected. Upon receiving this, the thermostat immediately changes the temperature setting to 23°C for the next hour. At the same time, the user preference is re-estimated based on the new observations, such as Fig.12 Then, based on the re-estimated user preferences and the updated electricity price and outdoor temperature forecast for this period, the 11:36-22:36 ( Fig.11 The purple line in the figure shows the new temperature setting schedule.

[0138] In this way, the thermostat cycles between condition (1) and condition (2), so the temperature setting schedule will be updated at least once every hour.

[0139] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

[0140] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

Claims

1. A method for learning user preferences and generating operation strategies for air-conditioning loads in a distribution network, characterized in that: The following steps are involved: Step 1) Estimate user preferences based on quantile regression method: By estimating the temperature set point T set The probability relationship between the user's preference and the electricity price e is used to estimate the user's preference; first, the user's total cost function C all Defined as the purchase cost C pur (C pur =eq) and discomfort level C dis The sum of T ideal The discomfort function C is constructed using the sixth-order Taylor polynomial as the center dis w=(w1,w2,...,w6) is the discomfort coefficient; Then the temperature setting T is expressed by the thermodynamic equation of the building. set Relationship with power consumption q: Where m is the order of the Taylor polynomial, w m is the mth component of the discomfort coefficient w, λ is the building heat dissipation coefficient, β is the thermal parameter of the building, Δt is the time interval, T0 is the initial indoor temperature, q is the power consumption, T out is the outdoor temperature; the optimal solution of the above formula should satisfy the Karush-Kuhn-Tucher (KKT) condition, and further obtain: Then, based on the above formula (2), the user's temperature setting set and electricity price e, we get the joint probability density distribution of w = (w1, w2, ..., w6); When more evidence or information is available, quantile regression is applied to update the probability of the hypothesis, and the probability distribution of the discomfort coefficient w is obtained by quantile regression to estimate the temperature set point T set The probability relationship between the electricity price e is as follows: Define a prior distribution to express the "degree of belief" in the possible values ​​of w: In the formula, α is a hyperparameter, the prior and likelihood are both Gaussian distributions, so the posterior is also Gaussian distribution: p(w|e ob ,a,s 2 )=N(μ w ,∑) (4) In the formula, e ob is the electricity price vector, σ 2 is the Gaussian noise ε n The variance of w The mean of the posterior vector is the ∑ posterior covariance matrix, defining the matrix Φ nm =(T n -T ideal ) m Then we have: Choose α=α MP and in is the marginal likelihood p(e ob |α,σ 2 )’s maximum value: Where I is the unit matrix and N is the number of observations. This maximization problem is solved by particle swarm optimization. The optimal parameters are found through Occam’s razor, and then the joint probability density distribution of w is obtained. In summary, we can get: Where e est It is T set.est Estimation of the electricity price, μ e is the mean of the marginal distribution, is the variance of the marginal distribution, and the other parameters are the same as above; according to formula (7), we can estimate T set The probability relationship between it and the electricity price e; Step 2), generate the optimal temperature setting plan based on the random expectation model: When the discomfort coefficient w is determined, the optimization problem in equation (1) is solved to generate the optimal temperature setting schedule. In order to obtain a more practical and comprehensive schedule, a single cycle is used to further discuss equation (1): Where T sche is the temperature setting table, T min is the lower limit of the temperature setting range, T max is the upper limit of the allowable range of temperature setting, and the other parameters are the same as above. In order to further reduce the electricity bill by utilizing the heat storage capacity of the building under the real-time electricity price, the multi-period random expectation model of formula (1) is established in formula (9): Where e t is the electricity price forecast at time t, q t is the power consumption at time t, q max is the rated capacity of the high voltage AC system, T sche.t is the temperature setting schedule at time t, T out.t is the outdoor temperature forecast at time t, T max It is the upper limit of the allowable range of temperature setting, and the other parameters are the same as above; Then, the Monte Carlo simulation is used to solve equation (9). In time period t, equation (9) is solved to obtain the temperature setting plans for time period t+1, time period t+2, ..., time period t+T, which will minimize the expected total cost of T time periods. Since the multi-period random expectation model can produce a more economical temperature setting plan than the single-period random expectation model, the thermostat operation strategy proposed in the subsequent step 3 will be based on equation (9); Step 3), a method for generating a thermostat automatic operation strategy based on user preference estimation and temperature setting scheduling: Based on the research of step 1 user preference estimation and step 2 temperature setting scheduling, a method for generating thermostat automatic operation strategy is designed, and the operation strategy is divided into three stages: startup stage, user preference estimation stage and automatic operation stage. In the startup phase, users specify their ideal temperature and select a default mode to start the thermostat. In the user preference estimation phase, the thermostat estimates user preferences based on user behavior observations and generates a temperature setting schedule on its own when appropriate. In the automatic operation stage, the outdoor temperature and real-time electricity price are first predicted for the next T hours. The random expectation model is solved through Monte Carlo simulation to generate a temperature setting plan for the next T hours. The thermostat will execute the schedule in the next hour while continuing to detect user input. If no new user input is detected in the next hour, the thermostat will repeat the process of this stage.

Images