An intelligent household appliance electricity daily load prediction method

Abstract

The application discloses a kind of intelligent household appliance electricity daily load prediction methods, comprising: obtaining 96 point history load data of intelligent household appliance electricity, meteorological data, date data, and pre-processing;According to season, date data, the history load data after pre-processing is sequentially clustered;Using Pearson correlation analysis method, quantitatively find out the meteorological factor of big correlation with intelligent household appliance electricity daily load as data sample input prediction pre-reference data;Particle swarm optimization (PSO) algorithm is used to optimize the parameters of LSSVM model;Residual correction is carried out on the optimized LSSVM model using BP neural network;The intelligent household appliance electricity daily load to be predicted is predicted.The method considers multiple influence elements, uses combination mode to complement each other, improves the prediction accuracy;Intelligent household appliance electricity daily load data can be accurately predicted, which provides basis for the safe operation of power grid and power planning, and can provide decision support for demand side management.

Description

Technical Field

[0001] This invention relates to the field of smart home appliance power consumption prediction technology, and in particular to a method for predicting the daily load of smart home appliances. Background Technology

[0002] Currently, existing load forecasting mainly focuses on macro-level forecasting of large-scale regional power grids, without conducting fine-grained forecasting of individual residential users' electricity loads, let alone fine-grained load forecasting of individual smart home appliances for residential users.

[0003] However, with rapid economic development and the continuous improvement of residents' living standards, various high-end and mid-range household appliances have entered people's lives, and residential electricity consumption accounts for a considerable proportion of total social electricity consumption. The continuous development of society has led to an exponential growth trend in residential electricity consumption, and the peak-valley difference in the power grid is gradually widening. Residential electricity load forecasting is the basis for conducting power balance analysis.

[0004] As the level of power grid intelligence continues to improve, macro-level forecasting results can no longer meet the needs of power grid development. Therefore, it is necessary to conduct more granular load forecasting for individual smart home appliances of residential users.

[0005] However, current methods for predicting the load of individual smart home appliances for residential users often employ single or combined methods. In the early stages, the training samples are limited, and as the sample size increases, the accuracy, convergence, and stability of the prediction gradually decrease, making it difficult to obtain ideal prediction results.

[0006] Therefore, how to accurately predict residential electricity load is an urgent problem to be solved in the development of smart grids. Summary of the Invention

[0007] The purpose of this invention is to provide a method for predicting the daily load of electricity consumption for smart home appliances that at least partially solves the above-mentioned technical problems. This method is based on the consideration of multiple influencing factors, so as to achieve scientific, reasonable and accurate prediction of electricity consumption for residential home appliances and improve prediction accuracy.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0009] This invention provides a method for predicting the daily electricity load of smart home appliances, comprising the following steps:

[0010] S10. Obtain historical load data, meteorological data, and date data for 96 points of smart home appliance power consumption, and perform preprocessing.

[0011] S20. Cluster the preprocessed historical load data according to seasonal and date data in sequence;

[0012] S30. Based on the clustering results, Pearson correlation analysis is used to quantitatively identify meteorological factors that are highly correlated with the daily load of smart home appliances as data sample inputs for prediction reference data.

[0013] S40. Based on a portion of the preprocessed data as the first training data, an LSSVM model is constructed for training, and the Particle Swarm Optimization (PSO) algorithm is used to optimize the parameters of the LSSVM model.

[0014] S50. Based on a portion of the preprocessed data as the second training data, a BP neural network is used to correct the residuals of the optimized LSSVM model.

[0015] S60. Based on the optimized LSSVM model and BP neural network model, predict the daily electricity load of smart home appliances that need to be predicted.

[0016] Further, in step S10, the preprocessing of historical load data from 96 points of smart home appliance power consumption includes:

[0017] If the first point of the daily load curve is not positive, it is replaced by the first positive point thereafter; if any other point of the daily load curve is not positive, it is replaced by the previous positive point.

[0018] The average value of two adjacent points is used to correct spurious data points; spurious data points refer to the load at a certain moment compared with the load values ​​before and after it, if the deviation is more than a preset multiple.

[0019] Historical load data is normalized using logarithms to ensure that the processed load falls within a common range of numerical characteristics.

[0020] Further, in step S10, the meteorological data preprocessing corresponding to the 96 historical load data points of smart home appliance electricity consumption includes:

[0021] The data sequences of temperature, humidity, wind force and precipitation are normalized so that all processed data sequences are evenly distributed in the range [0,1].

[0022] Further, step S20 includes:

[0023] The preprocessed historical load data was first clustered according to the four seasons of spring, summer, autumn and winter.

[0024] Then, based on the date, a second clustering was performed on the historical load data for spring and autumn, and summer and winter, to obtain the clustering results for weekdays, rest days, and holidays.

[0025] Furthermore, in step S30, the highest temperature, lowest temperature, and corresponding humidity in the meteorological data corresponding to the historical load data are selected as data sample input prediction reference data.

[0026] Further, in step S40, the parameters of the LSSVM model are optimized using the Particle Swarm Optimization (PSO) algorithm, including:

[0027] 1) Initialize various parameters of the Particle Swarm Optimization (PSO) algorithm, including: swarm size, learning factor, maximum number of iterations, initial position and velocity of particles;

[0028] 2) Use the LSSVM model corresponding to each particle vector to predict the learning samples, obtain the prediction error of the current position value of each particle, and use it as the fitness value of each particle; then compare the current fitness value of each particle with its own best fitness value. If the latter is better, then the position corresponding to the best fitness value of the particle is taken as the best position of the particle.

[0029] 3) Compare the fitness value of each particle's own best position with the fitness value of the group's best position. If the fitness value of the best position of the particle is better, then the best position of the particle is taken as the best position of the group.

[0030] 4) Update the particle's velocity and position based on the calculated inertia weights;

[0031] 5) Check if the optimization termination condition is met. If it is met, end the optimization and find the optimal solution; otherwise, go to step 2) and continue a new round of search. The optimization termination condition is: reaching the preset maximum number of iterations or preset accuracy.

[0032] Further, step S50 includes:

[0033] Based on a portion of the preprocessed data as the second training data, the highest temperature, lowest temperature and corresponding humidity in the second training data are used as input values.

[0034] The deviation between the measured values ​​of the historical load data of the second training data and the predicted values ​​of the optimized LSSVM model is used as the output value.

[0035] The input and output values ​​are used to construct training samples, and a 3-layer feedforward BP neural network model is established.

[0036] Furthermore, the process of establishing the BP neural network model includes:

[0037] ① Neural network initialization: Provide the initial weights W between the input layer and the hidden layers. ij Initial threshold b jAnd the initial weights W between the hidden layer and the output layer. jk Initial threshold b m And set the initial learning rate η, the termination error e, and the maximum number of iterations λ;

[0038] ② Forward signal transmission: Output of the j-th neuron in the hidden layer:

[0039]

[0040] In the formula, f is the activation function of the hidden layer, and the mapping function of the hidden layer neurons adopts the Sigmoid function; x i Let be the input value of the i-th input node, n be the total number of input values, and l be the total number of neurons in the hidden layer.

[0041] Output of the k-th neuron in the output layer:

[0042]

[0043] In the formula, f is the activation function of the output layer, which is the Purelin function; m is the total number of neurons in the output layer; the learning error function is:

[0044]

[0045] In the formula, y k This is the actual output of the sample;

[0046] ③ Error backpropagation and weight threshold training and update: The connection weights from the j-th neuron in the hidden layer to the k-th neuron in the output layer have the following update increment calculation formula:

[0047]

[0048] Similarly, the threshold update increment can be obtained:

[0049]

[0050] The formula for calculating the update increment of the connection weights and thresholds between the i-th neuron in the input layer and the j-th neuron in the hidden layer is as follows:

[0051]

[0052]

[0053] The neuron connection weights and thresholds for the next round of network learning and training are iteratively updated based on the incremental changes in the obtained neuron connection weights and thresholds of each layer, where t is the current iteration number.

[0054] w ij (t+1)=w ij(t)+Δw ij (t)

[0055] w jk (t+1)=w jk (t)+Δw jk (t)

[0056] b j (t+1)=b j (t)+Δb j (t)

[0057] b k (t+1)=b k (t)+Δb k (t)

[0058] ④ Determine if learning has ended: If the error is less than or equal to e or the number of iterations reaches the specified value λ, then learning ends and the predicted value of the learning sample and the relative error between the predicted value and the actual value APE are output.

[0059]

[0060] Alternatively, return to step ③ and continue training;

[0061] ⑤ Establish a BP neural network residual model to obtain the mapping relationship model between the highest temperature, lowest temperature, humidity and the prediction error of the PSO-LSSVM model, i.e., the BP neural network residual model.

[0062] Compared with the prior art, the present invention has the following beneficial effects:

[0063] A method for predicting daily load of smart home appliance electricity consumption includes: acquiring historical load data, meteorological data, and date data for 96 points of smart home appliance electricity consumption, and preprocessing them; clustering the preprocessed historical load data according to season and date data; based on the clustering results, using Pearson correlation analysis, quantitatively identifying meteorological factors with high correlation to the daily load of smart home appliance electricity consumption as input reference data for prediction; constructing an LSSVM model for training based on a portion of the preprocessed data as the first training data, and optimizing the parameters of the LSSVM model using the Particle Swarm Optimization (PSO) algorithm; using a portion of the preprocessed data as the second training data, and using a Backpropagation (BP) neural network to correct the residuals of the optimized LSSVM model; and predicting the daily load of smart home appliance electricity consumption based on the optimized LSSVM model and the BP neural network model. This method considers multiple influencing factors and uses a combination approach to leverage the strengths and compensate for the weaknesses of the selected methods, improving prediction accuracy; it can accurately predict the daily load data of smart home appliance electricity consumption, providing a basis for the safe operation of the power grid and power planning, and also providing decision support for demand-side management. Attached Figure Description

[0064] Figure 1 A flowchart of the daily load prediction method for smart home appliances provided in an embodiment of the present invention;

[0065] Figure 2 This is a schematic diagram illustrating the overall principle of the smart home appliance daily load prediction method provided in this embodiment of the invention. Detailed Implementation

[0066] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0067] In the description of this invention, it should be noted that the terms "upper," "lower," "inner," "outer," "front end," "rear end," "both ends," "one end," and "the other end," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0068] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installed," "equipped with," "connected," etc., should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be a connection within two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0069] Reference Figure 1 As shown, the present invention provides a method for predicting the daily electricity load of smart home appliances, comprising:

[0070] S10. Obtain historical load data, meteorological data, and date data for 96 points of smart home appliance power consumption, and perform preprocessing.

[0071] S20. Cluster the preprocessed historical load data according to seasonal and date data in sequence;

[0072] S30. Based on the clustering results, Pearson correlation analysis is used to quantitatively identify meteorological factors that are highly correlated with the daily load of smart home appliances as data sample inputs for prediction reference data.

[0073] S40. Based on a portion of the preprocessed data as the first training data, an LSSVM model is constructed for training, and the Particle Swarm Optimization (PSO) algorithm is used to optimize the parameters of the LSSVM model.

[0074] S50. Based on a portion of the preprocessed data as the second training data, a BP neural network is used to correct the residuals of the optimized LSSVM model.

[0075] S60. Based on the optimized LSSVM model and BP neural network model, predict the daily electricity load of smart home appliances that need to be predicted.

[0076] In this embodiment, the historical load data, meteorological data, and date data of 96 points of smart home appliance electricity consumption are first preprocessed; the preprocessed historical load data are then clustered according to season and date data; based on the clustering results, Pearson correlation analysis is used to quantitatively identify meteorological factors with high correlation to the daily load of smart home appliance electricity consumption as input reference data for prediction; the particle swarm optimization (PSO) algorithm is used to optimize the penalty coefficient C and radial basis function parameter б of the LSSVM (least squares support vector machine), and a backpropagation (BP) neural network is used to correct the residuals of the model. Figure 2 As shown, the specific process is as follows:

[0077] (1) Data Acquisition

[0078] We obtained historical daily load data for 96 smart home appliances from the electricity consumption information collection system, and compiled historical meteorological data and date data to prepare for the next step.

[0079] (2) Abnormal data preprocessing

[0080] ① Preprocessing of non-positive points

[0081] If load data is not read or is read incorrectly, non-positive values ​​will appear in the load data. The method to handle this type of data is as follows: if the first point of the load curve for the day is non-positive, replace it with the first positive point after that point; if other points of the load curve for the day are non-positive, replace them with the previous positive point.

[0082] ② Preprocessing of spurious data points

[0083] The load at a certain moment is compared with the load values ​​before and after it. If the deviation is more than three times, it is considered a false data point. The method for processing such data is to use the average value of two adjacent points to correct the false data point.

[0084] (3) Load data normalization processing

[0085] To simplify calculations and reduce the magnitude, logarithmic processing is applied to the load data so that the processed load falls within a common range of numerical characteristics.

[0086] (4) Meteorological data preprocessing

[0087] To ensure that the acquired meteorological data, such as temperature, humidity, wind speed, and precipitation, have dimensionless quantities of the same scale, a normalization formula was used to process each data sequence. All processed data sequences are uniformly distributed within the range [0, 1]. The normalization formula is as follows:

[0088]

[0089] Among them, T' i For a certain meteorological factor after normalization, T i For the original value, T min Let T be the minimum value in the sequence. max Let n be the maximum value in the sequence, and n be the length of the sequence.

[0090] (5) Determination of similar days

[0091] Historical load data were clustered according to the four seasons of spring, summer, autumn and winter. Based on the clustering results, the historical load data were divided into two categories: spring / autumn and summer / winter. The historical load data for spring / autumn and summer / winter were further clustered according to date, and each category was further subdivided into weekdays, ordinary rest days (non-statutory holidays), and public holidays.

[0092] (6) Correlation analysis of meteorological factors

[0093] There are many factors that affect electricity load. In order to find out the impact of different meteorological factors on electricity load, Pearson correlation analysis was used to calculate the correlation between historical load data and the collected meteorological factors. Finally, the highest temperature, lowest temperature and humidity were selected as data samples to be used as the pre-prediction reference data.

[0094] (7) Establishment of PSO-LSSVM model

[0095] The preprocessed historical load data is divided into two parts, for example, 90% for training data and 10% for validation data. The training data is then divided into two equal parts as two sets of data. The first set is used to build the PSO-LSSVM model, and the second set is used to build the BP residual model.

[0096] To model using LSSVM, the first step is to determine the radial basis function kernel parameter б and the penalty coefficient C. The radial basis function kernel parameter б quantifies the correlation between support vectors. If б is set too large, it can lead to undertraining of the LSSVM; if it is set too small, there is a risk of overtraining. Similarly, a large penalty coefficient C indicates a small penalty for bias, which can easily lead to undertraining. Conversely, a small penalty indicates a large penalty for bias, requiring consideration of more points and potentially leading to overtraining. In this embodiment, the PSO algorithm is used to optimize the parameters C and б of the LSSVM. The optimal parameters are then used as the LSSVM parameters to establish the PSO-LSSVM model. The specific process is as follows:

[0097] ① PSO-based LSSVM parameter optimization:

[0098] 1) Initialize various parameters of PSO: population size, learning factor, maximum number of iterations, initial position and velocity of particles, etc.

[0099] 2) Use the LSSVM corresponding to each particle vector to predict the learning samples, obtain the prediction error of the current position value of each particle, and use it as the fitness value of each particle. Then compare the current fitness value of each particle with its own best fitness value. If it is better, then the current position of the particle is taken as the best position of the particle.

[0100] 3) Compare the fitness value of each particle's own optimal position with the fitness value of the group's optimal position. If the fitness value of the particle is better, then the optimal position of the particle is taken as the optimal position of the group.

[0101] 4) Update the particle's velocity and position based on the calculated inertia weights.

[0102] 5) Check if the optimization termination condition is met (reaching the preset maximum number of iterations or preset precision). If it is met, end the optimization and find the optimal solution; otherwise, go to step 2) and continue a new round of search.

[0103] ②Establishment of the PSO-LSSVM model

[0104] The optimal parameters obtained from PSO are then substituted into the LSSVM model to establish the PSO-LSSVM model.

[0105] (8) Establishment of the BP residual correction model

[0106] The PSO-LSSVM model was used to predict the second set of training data, and the predicted values ​​were recorded. The highest temperature, lowest temperature, and humidity in the second set of data were used as input values ​​X. i The deviation between the measured value and the PSO-LSSVM predicted value is used as the output value T. iBuild training samples (X) i T i A 3-layer feedforward BP neural network is established. The specific process is as follows:

[0107] ① Neural network initialization: Provide the initial weights W between the input layer and the hidden layers. ij and initial threshold b j and the initial weights W between the hidden layer and the output layer jk and initial threshold b m The initial learning rate η, the termination error e, and the maximum number of iterations λ are set.

[0108] ② Forward signal transmission: Output of the j-th neuron in the hidden layer:

[0109]

[0110] In the formula, f is the activation function of the hidden layer, and the Sigmoid function is often used as the mapping function for hidden layer neurons; x i Let be the input value of the i-th input node, n be the total number of input values, and l be the total number of neurons in the hidden layer.

[0111] Output of the k-th neuron in the output layer:

[0112]

[0113] In the formula, f is the activation function of the output layer, which is generally the Purelin function; m is the total number of neurons in the output layer;

[0114] Learning error function:

[0115]

[0116] In the formula, y k This is the actual output of the sample.

[0117] ③ Error backpropagation and weight threshold training and update. The connection weights from the j-th neuron in the hidden layer to the k-th neuron in the output layer are calculated using the following formula for the update increment:

[0118]

[0119] Similarly, the threshold update increment can be obtained:

[0120]

[0121] The formula for calculating the update increment of the connection weights and thresholds between the i-th neuron in the input layer and the j-th neuron in the hidden layer is as follows:

[0122]

[0123]

[0124] The neuron connection weights and thresholds for the next round of network learning and training are iteratively updated based on the incremental changes in the obtained neuron connection weights and thresholds of each layer, where t is the current iteration number.

[0125] w ij (t+1)=w ij (t)+Δw ij (t)

[0126] w jk (t+1)=w jk (t)+Δw jk (t)

[0127] b j (t+1)=b j (t)+Δb j (t)

[0128] b k (t+1)=b k (t)+Δb k (t) (8)

[0130] ④ Determine if learning has ended. If the error is less than or equal to e or the number of iterations reaches the specified value λ, then learning ends and the predicted value of the learning sample and the relative error between the predicted value and the actual value (APE) are output.

[0131]

[0132] In the formula, y(t) is the actual value and o(t) is the predicted value; otherwise, return to step ③ to continue training.

[0133] ⑤ Establish a BP neural network residual model to obtain the mapping relationship model between the highest temperature, lowest temperature, humidity and the prediction error of the PSO-LSSVM model, i.e., the BP neural network residual model.

[0134] (9) Establishment of the PSO-LSSVM-BP combined model

[0135] Historical electricity consumption data of smart home appliances to be predicted is input into the PSO-LSSVM model and the BP neural network residual model. The load forecast value is obtained from the PSO-LSSVM model, and the residual forecast value is obtained from the BP residual model. In other words, the load forecast value obtained from the PSO-LSSVM model and the residual forecast value obtained from the BP residual model are added together to obtain the final load forecast value. The two models are combined to obtain the PSO-LSSVM-BP combined model.

[0136] (10) Validation of the PSO-LSSVM-BP combined model

[0137] The validation data was input into the PSO-LSSVM-BP combined model to verify its predictive performance. Experiments demonstrate that the combined model exhibits excellent data fitting ability and extremely high prediction accuracy.

[0138] The method for predicting daily electricity load for smart home appliances provided in this invention has the following advantages:

[0139] 1. Iteratively group and cluster historical daily load data, dividing the historical load data into two categories: spring and autumn, and summer and winter. Cluster the historical load data of spring and autumn and summer and winter according to the date. Further subdivide each category into weekdays, ordinary rest days (non-statutory holidays), and holidays, and establish similar day division intervals to provide more accurate data support for daily load forecasting.

[0140] 2. To address the blind selection problem in LSSVM parameters, the PSO algorithm is used to optimize the LSSVM parameters C and б. To address the issue that the PSO algorithm cannot continue optimization after reaching a certain convergence point in the later stages, a BP neural network residual correction model is established, leveraging the high reliability and good fault tolerance of the BP neural network.

[0141] 3. By effectively utilizing the historical daily load data and dividing it into training data and validation data, a model was established and its feasibility was verified.

[0142] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method of predicting daily power consumption of a smart home appliance, characterized by, The method comprises the following steps: S10, obtaining 96-point historical load data, meteorological data and date data of the intelligent household appliance, and preprocessing the data; S20, clustering the preprocessed historical load data according to seasons and date data; S30, based on the clustering result, using the Pearson correlation analysis method to quantitatively find out meteorological factors with high correlation with the daily load of the intelligent household appliance as data sample input for the prediction pre-reference data; S40, based on part of the preprocessed data as first training data, constructing an LSSVM model for training, and using the particle swarm optimization (PSO) algorithm to optimize the parameters of the LSSVM model; S50, based on part of the preprocessed data as second training data, using a BP neural network to correct the residual error of the optimized LSSVM model; S60, predicting the daily load of the intelligent household appliance according to the optimized LSSVM model and the BP neural network model; In the step S30, the highest temperature, the lowest temperature and the corresponding humidity in the meteorological data corresponding to the historical load data are selected as the data sample input for the prediction pre-reference data; In the step S40, the particle swarm optimization (PSO) algorithm is used to optimize the parameters of the LSSVM model, which comprises: 1) initializing various parameters of the particle swarm optimization (PSO) algorithm, including the population size, the learning factor, the maximum number of iterations, the initial position and speed of the particles; 2) predicting the learning sample by using the LSSVM model corresponding to each particle vector to obtain the prediction error of the current position value of each particle, and taking the prediction error as the fitness value of each particle; then comparing the current fitness value of each particle with the optimal fitness value of the particle itself, and if the latter is better, taking the position corresponding to the optimal fitness value of the particle itself as the optimal position of the particle; 3) comparing the optimal position fitness value of each particle with the fitness value of the optimal position of the population, and if the former is better, taking the optimal position of the particle as the optimal position of the population; 4) updating the speed and position of the particle according to the calculated inertia weight; 5) checking whether the optimization ending condition is met, and if yes, ending the optimization and obtaining the optimal solution; otherwise, going to step 2) to continue the new round of search; the optimization ending condition is that the maximum number of iterations or the preset accuracy is reached; The step S50 comprises: based on part of the preprocessed data as second training data, taking the highest temperature, the lowest temperature and the corresponding humidity in the second training data as input values; taking the deviation between the measured value of the historical load data of the second training data and the prediction value of the optimized LSSVM model as an output value; establishing a 3-layer feedforward BP neural network model by using the input values and the output values to form training samples; The step S60 comprises: inputting the intelligent household appliance load history data to be predicted into the PSO-LSSVM model and the BP neural network residual error model; obtaining the load prediction value from the PSO-LSSVM model and the residual error prediction value from the BP residual error model; The load prediction value and the residual prediction value are added together as a final load prediction value.

2. The method of claim 1, wherein the method comprises: In the step S10, the preprocessing of the 96-point historical load data of the intelligent household appliance includes: When the first point of the daily load curve is not positive, the first positive point thereafter is used to replace it; when other points of the daily load curve are not positive, the previous positive point of the point is used to replace it; The average value of the adjacent two points is used to correct the false data point; the false data point refers to the comparison of the load at a certain time and the load before and after it, if the deviation is outside the preset multiple; The logarithm is used to normalize the historical load data, so that the processed load is within the common numerical characteristic range.

3. The method of claim 1, wherein the method comprises: In the step S10, the preprocessing of the 96-point historical load data of the intelligent household appliance includes: The data sequences of temperature, humidity, wind power and precipitation are normalized, so that all the processed data sequences are uniformly dispersed in the range of [0, 1].

4. The method of claim 1, wherein the method comprises: The step S20 includes: According to the four seasons of spring, summer, autumn and winter, the historical load data after preprocessing is clustered for the first time; According to the date, the historical load data of the spring-autumn and summer-winter types is clustered for the second time, and the type clustering results of working days, rest days and holidays are obtained.

5. The method of claim 1, wherein the method comprises: The establishment process of the BP neural network model includes: ① Initialization of neural network: initial weights W ij , initial thresholds b j between input layer and hidden layer are given, and initial weights W jk , initial thresholds b m between hidden layer and output layer are given, and initial learning rate η, termination error e and maximum iteration number λ are set; The output of the jth neuron in the hidden layer: In the formula, f is the excitation function of the hidden layer, and the mapping function of the hidden layer neuron adopts the Sigmoid function; x i is the input value of the i th input node, n is the total number of input values, and l is the total number of hidden layer neurons. The output of the kth neuron in the output layer: In the formula, f is the excitation function of the output layer, which is the Purelin function; m is the total number of neurons in the output layer; The learning error function: In the formula, y k is the actual output of the sample; The connection weight value from the jth neuron in the hidden layer to the kth neuron in the output layer has the following update increment calculation formula: Similarly, the threshold value update increment is obtained: The calculation formula of the update increment of the connection weight value and the threshold value between the ith neuron in the input layer and the jth neuron in the hidden layer is: According to the obtained change increment of the neural connection weight value and the threshold value of each layer, the neural connection weight value and the threshold value for the next round of network learning and training are iteratively updated, t is the current iteration number, that is: w ij (t+1) = w ij (t) + Δw ij (t) w jk (t+1) = w jk (t) + Δw jk (t) b j (t+1) = b j (t) + Δb j (t) b k (t+1) = b k (t) + Δb k (t) If the error is less than or equal to e or the iteration number reaches the specified value λ, the learning is ended and the prediction value of the learning sample and the relative error APE between the prediction value and the actual value are outputted; Otherwise, return to step ③ for continuous training; ⑤ Establish the BP neural network residual error model to obtain the mapping relationship model of the highest temperature, the lowest temperature, the humidity and the PSO-LSSVM model prediction error, that is, the BP neural network residual error model.

Images