An enhanced residential load forecasting method based on event response knowledge guidance

Abstract

The application discloses an enhanced residential load prediction method based on event response knowledge guidance, and comprises the following steps: acquiring power sequence data; performing state identification on each power sequence data; training a first neural network to obtain a power use event prediction module; taking a historical time and power data corresponding to the historical time as input, taking a target time and power data corresponding to the target time as output, training a second neural network, and using a state probability corresponding to a target time state label output by the power use event prediction module, a prediction output of the second neural network and a second loss function calculated by an actual target value to adjust a training process of the second neural network, so as to obtain a power prediction model; and outputting enhanced residential load prediction data through the power prediction model. The application can improve the prediction accuracy of the power prediction model, and further improve the accuracy of the enhanced residential load prediction data.

Description

Technical Field

[0001] This invention belongs to the field of power forecasting and relates to a forecasting method, specifically an enhanced residential load forecasting method based on event response knowledge guidance. Background Technology

[0002] Residential load forecasting (RLF) aims to predict future electricity consumption by individual consumers, reflecting their anticipated household electricity demand and corresponding consumption behavior. Within the power system, various participants can benefit from predictive analysis of electricity use. For system operators, RLF provides the total future electricity demand of residents within a specific area. Based on current power resource conditions and demand, operators will develop corresponding power dispatching or demand response strategies. For example, by adopting different time-of-use (TOU) pricing schemes, consumers can be incentivized to dynamically adjust their electricity usage to achieve peak load reduction and valley filling. For consumers, residential load forecasting enables better household dispatching to cope with real-time pricing (RTP) or to pre-determine energy storage. Accurate residential load forecasting contributes to the optimal allocation of power resources, improving the efficiency and reliability of the power system. Therefore, exploring accurate residential load forecasting is becoming a popular research area in both industry and academia.

[0003] The deployment of smart meters and advancements in non-intrusive load monitoring (NILM) technology have provided a pathway to acquiring large amounts of fine-grained data, laying the foundation for accurate analysis of residential user behavior. Based on this, many studies have proposed methods for learning the inherent dependencies in large-scale historical data, namely the seasonality and growth patterns of historical data. Most existing RLF methods can generally be categorized into two levels based on different forecasting levels: household level and device level. In effect, RLF at both levels constitutes a multivariate time series forecasting (MTSF) problem. This involves identifying historical patterns of internal variables and modeling the relationships between variables to reflect consumer electricity usage. However, electricity usage patterns at the device level and household level are often different, and the relationships between them are also dynamically changing.

[0004] Recent advances in multivariate time series forecasting (MTSF) have yielded advanced models adept at managing complex patterns of internal and intervariate variables, particularly those categorized as Transformer-based and Multilayer Perceptron (MLP)-based models. Transformer-based models capture long-term dependencies through self-attention mechanisms and use positional encoding techniques to preserve temporal information. MLP-based models, on the other hand, retain complete temporal information through a default linear architecture. However, residential load series are typically sparse and noisy. This is because load series mix actual consumer behavior with noise, and changes in actual behavior are sparse relative to noise. Existing forecasting methods attempt to extract inherent dependencies directly from historical and future load records. Ignoring the mining of sparse knowledge from dense time series leads existing methods to learn not only the corresponding sequence patterns of electricity consumption behavior but also the sequence patterns of noise when modeling household load forecasting. Therefore, these methods can be susceptible to noise and interference. Summary of the Invention

[0005] To address the aforementioned shortcomings in existing technologies, this invention provides an enhanced residential load forecasting method based on event response knowledge, which solves the problem of inaccurate electricity usage forecasting in existing technologies.

[0006] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows:

[0007] An enhanced residential load forecasting method guided by event response knowledge includes the following steps:

[0008] S1. Obtain power sequence data. The historical power sequence includes historical time, power data corresponding to the historical time, target time, and power data corresponding to the target time.

[0009] S2. Perform state identification on each power sequence data, and obtain the historical time state label corresponding to each historical time and the target time state label corresponding to each target time.

[0010] S3. Take the historical time and the corresponding power data as input, and the target time state label as output, train the first neural network to obtain the power usage event prediction module. The loss function of the power usage event prediction module is the first loss function.

[0011] S4. Take the historical time and the corresponding power data as input, and the target time and the corresponding power data as output to train the second neural network. Then, use the state probability corresponding to the target time state label output by the power usage event prediction module, the prediction output of the second neural network, and the second loss function calculated from the actual target value to adjust the training process of the second neural network to obtain the power prediction model.

[0012] S5. Input the predicted power sequence data into the power prediction model, and output the enhanced residential load prediction data through the power prediction model.

[0013] The beneficial effects of the above scheme are:

[0014] (1) This invention uses historical load sequences to extract knowledge related to events and estimate changes in the operating status of electrical appliances to mark the beginning and end of events. Compared with directly learning the relationship between historical load values ​​and future load values ​​to be predicted, it pays more attention to the load data corresponding to power usage events and can effectively extract event-related knowledge that is beneficial to load prediction.

[0015] (2) This invention encourages residential load prediction models to pay more attention to sparse event information and reduce the impact of noise through an event-response-based knowledge guidance mechanism. Without changing the structure and reasoning process of the original prediction model, it improves the robustness of the power prediction model and ensures the prediction accuracy of the power prediction model.

[0016] Furthermore, in step S2, the historical time status label and the target time status label include: ON, OFF, and high performance (HP), respectively.

[0017] Further, step S2 includes:

[0018] A state recognition algorithm is used to determine the historical state label for each historical time and the target state label for each target time, respectively, based on the power data corresponding to the historical time and the power data corresponding to the target time.

[0019] The beneficial effects of the above-mentioned further solutions are: probabilistic modeling and prediction based on power data at historical and target times can lead to more intelligent and accurate power forecasting and management, which helps to improve the efficiency, stability and reliability of the power system.

[0020] Further, step S4 includes:

[0021] S41. Determine the state probability based on the target time state label output by the power usage event prediction module;

[0022] S42. The state probability is used as the weight coefficient in the adjustment formula, and the second loss function calculated by combining the predicted output of the second neural network and the actual target value is used to adjust the training process of the second neural network.

[0023] The beneficial effects of the above-mentioned further scheme are: adjusting the output error of the load forecasting module based on state information and probability weights, and using the adjusted output error as a loss function to update the network parameters of the load forecasting module, can improve the accuracy and applicability of the power forecasting model, help better guide the operation and management of the power system, and achieve more intelligent and efficient power forecasting.

[0024] Furthermore, in step S4, the second loss function is:

[0025]

[0026] in, Indicates by z t The logical value specifies the error between the predicted and actual values ​​in the specified space, where y represents the actual value of the power series data. z represents the predicted value of power series data. t This represents the output of the last layer of the neural network in the power usage event prediction module.

[0027] Furthermore, in step S42, the adjustment formula is:

[0028]

[0029]

[0030]

[0031] in, Indicates by z t The error between the predicted and actual values ​​within the space specified by the logical value. This represents the general regression loss. This represents the final loss of the load forecasting module, where D represents the number of dimensional variables, t represents the time in the electricity sequence data, and H represents the number of forecast time steps. This represents the logitsZ value of the variable in the j-th dimension at the i-th time step. Let represent the predicted value of the variable in the j-th dimension at the i-th time step. This represents the actual value of the variable in the j-th dimension at the i-th time step. Y represents the predicted value of all time steps and variables, Y represents the actual value of all time steps and variables, and α is a hyperparameter.

[0032] The beneficial effects of the above-mentioned further solutions are: adjusting the model using the above formula can improve the model's flexibility, personalization, and performance optimization, enhance the model's transparency and real-time performance, thereby making the model more suitable for practical application scenarios, improving prediction results, and promoting the scientific and accurate nature of decision-making.

[0033] Furthermore, in step S2, the state recognition algorithm includes: k-means clustering algorithm.

[0034] The beneficial effects of the above-mentioned further scheme are: using the k-means clustering algorithm to identify the state of power forecast data can automatically discover the state in the data, provide interpretable labels, assist in the establishment and optimization of forecast models, and provide more information and support for the operation and decision-making of power systems.

[0035] Furthermore, in step S3, the first neural network includes: an input layer, a convolutional layer, a univariate feature extraction layer, a prediction layer, and an output layer.

[0036] Furthermore, in step S3, the first loss function is:

[0037]

[0038] Where D represents the number of dimensional variables, H represents the number of prediction time steps, and S τ,d This represents the actual state category of the d-th variable at time τ. This represents the probability that the d-th variable is in the actual state category at time τ.

[0039] The beneficial effects of the above-mentioned further solutions are: the use of loss functions can measure the difference between the model's prediction results and the true labels, optimize model training, detect whether the model is overfitting as early as possible, and take corresponding measures. Furthermore, loss functions can be used to balance the importance of different samples or targets, thereby more accurately measuring model performance. Attached Figure Description

[0040] Figure 1 This is a flowchart illustrating an enhanced residential load forecasting method guided by event response knowledge. Detailed Implementation

[0041] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0042] like Figure 1 As shown, an enhanced residential load forecasting method based on event response knowledge specifically includes the following steps:

[0043] S1. Obtain power sequence data. The historical power sequence includes historical time, power data corresponding to the historical time, target time, and power data corresponding to the target time.

[0044] In step S1, the power sequence data can be obtained from a power database or from a power data storage server. No specific restrictions are placed on the method or channel for obtaining the power sequence data.

[0045] S2. Perform state identification on each power sequence data, and obtain the historical time state label corresponding to each historical time and the target time state label corresponding to each target time.

[0046] In this embodiment, the historical time status label and the target time status label include: ON, OFF, and high performance (HP), which can be used to indicate the operating status of electrical appliances.

[0047] In this embodiment, step S2 may specifically include: using a state recognition algorithm on the power data corresponding to the historical time and the power data corresponding to the target time to determine the historical time state label corresponding to each historical time and the target time state label corresponding to each target time.

[0048] In this embodiment, the state recognition algorithm used for the power data corresponding to the historical time and the power data corresponding to the target time may include: k-means clustering algorithm.

[0049] For example, a washing machine can have three operating states: 0 represents OFF, 1 represents ON, and 2 represents high performance (HP).

[0050] In this embodiment, the process of obtaining state labels using the k-means clustering algorithm is as follows: the power sequence data is classified into different operating states using the k-means clustering algorithm to form basic labels S∈R. l×D The number of running state categories can be represented as N. 0:D For example, given the original load sequence (i.e., power sequence data) E∈R l×D Where l represents the time step and D represents the number of dimensional variables. Samples can be generated using a sliding window of size w, with the clustering state bounds set to mins and maxs. The k-means clustering algorithm is applied to each dimensional variable for clustering, and the optimal number of clusters is determined based on the contour score, serving as the final number of categories for each dimensional variable's running state.

[0051] S3. Using historical time and corresponding power data as input, and the target time state label as output, train the first neural network to obtain the power usage event prediction module. The loss function of the power usage event prediction module is the first loss function.

[0052] The first loss function is:

[0053]

[0054] Where D represents the number of dimensional variables, H represents the number of prediction time steps, and S τ,d This represents the actual state category of the d-th variable at time τ. This represents the probability that the d-th variable is in the actual state category at time τ.

[0055] for The softmax function can be used to calculate this, and the formula is as follows:

[0056]

[0057] Among them, z τ,d,c This represents the logical value of the d-th variable at time τ, which is in state class c. Indicates z τ,d,c Exponentiation.

[0058] In step S3, the first neural network includes: an input layer, a convolutional layer, a univariate feature extraction layer, a prediction layer, and an output layer.

[0059] In some embodiments, when training the second neural network, the power sequence data used for training can be the load sequence X. t-M:t Its data length is M.

[0060] Load sequence X t-M:t Input can be taken from the input layer and enter the convolutional layer φ. sl Perform convolution processing, convolutional layer φ sl The convolutional layer can have two layers, and each layer can be considered as a feature extractor. Therefore, the convolutional layer φ sl It can include two connected feature extractors, which can process the load sequence X. t-M:t Extract the data from the data to obtain the initial data X. sl The expression for this step can be:

[0061] X sl =φ sl (X t-L:t )

[0062] Among them, X t-L:t Represents the load sequence, φsl Indicates convolution processing, X sl This represents the initial data.

[0063] The initial data is input into the univariate feature extraction layer, which can contain Q univariate extractors (UEs). The size of Q is related to the dimensionality of the power series data being processed, such as the number of electrical appliances. For example, when the dimensionality of the power series data is 1, Q is 1; when the dimensionality is 2, Q is 2, and so on. Each univariate extractor can contain three layers of convolutional neural networks. Each convolutional layer can be viewed as a feature extractor used to extract features from the initial data. Finally, the univariate feature extraction layer outputs intermediate data. The data shape can be Where H represents the prediction time step, x (i) Represents the i-th initial data X sl , where n represents the number of running states. The expression for this step can be:

[0064] For i∈0,…,Q-1

[0065] in, This represents the intermediate data output by the i-th univariate extractor. X represents the univariate extractor that processes the i-th initial data (i.e., the i-th of the Q univariate extractors). sl Represents the convolutional layer φ sl The output of .

[0066] intermediate data The input can be a prediction layer, which can contain a fully connected network to integrate Q sets of intermediate data obtained from Q univariate extractors. The final prediction layer can output predicted data Z, the shape of which can be... Where Q represents the number of univariate extractors, x (i) Represents the i-th initial data X sl , where n represents the number of running states, and R represents the data shape. The expression for this step can be:

[0067] Z = φ pl (Z u )

[0068] in, Represents i intermediate data The data is spliced ​​together, φ pl This represents the prediction layer in the electricity usage event prediction module.

[0069] Finally, the prediction layer processes the prediction data, takes the highest-scoring data as the prediction result, and outputs the final data. The expression for this step can be:

[0070] For i∈0,…,D-1

[0071] Among them, Z (i) H represents the predicted data corresponding to the i-th univariate extractor, which can also be regarded as the predicted data corresponding to the i-th load sequence (the data processed by the i-th univariate extractor is the i-th load sequence); H represents the prediction time step.

[0072] Final data It can be output directly from the output layer, or it can be processed in the output layer and transformed into the target data before being output. No specific restrictions are imposed here.

[0073] S4. Using historical time and corresponding power data as input, and target time and corresponding power data as output, train the second neural network. Adjust the training process of the second neural network using the state probability corresponding to the target time state label output by the power usage event prediction module, the prediction output of the second neural network, and the second loss function calculated from the actual target value, to obtain the power prediction model.

[0074] Specifically, step S4 includes:

[0075] S41. Determine the state probability based on the target time state label output by the power usage event prediction module.

[0076] S42. The state probability is used as the weight coefficient in the adjustment formula, and the second loss function calculated by combining the predicted output of the second neural network and the actual target value is used to adjust the training process of the second neural network.

[0077] The second loss function in step S4 is:

[0078]

[0079] in, Indicates by z t The logical value specifies the error between the predicted and actual values ​​in the specified space, where y represents the actual value of the power series data. z represents the predicted value of power series data. t This represents the output of the last layer of the neural network in the power usage event prediction module.

[0080] Specifically, the adjustment formula in step S42 is:

[0081]

[0082]

[0083]

[0084] in, Indicates by z t The error between the predicted and actual values ​​within the space specified by the logical value. This represents the general regression loss. This represents the final loss of the load forecasting module, where D represents the number of dimensional variables, t represents the time in the electricity sequence data, and H represents the number of forecast time steps. This represents the logitsZ value of the variable in the j-th dimension at the i-th time step. Let represent the predicted value of the variable in the j-th dimension at the i-th time step. This represents the actual value of the variable in the j-th dimension at the i-th time step. Y represents the predicted value of all time steps and variables, Y represents the actual value of all time steps and variables, and α is a hyperparameter.

[0085] S5. Input the predicted power sequence data into the power prediction model, and output the enhanced residential load prediction data through the power prediction model.

[0086] In this embodiment, by acquiring enhanced residential load forecasting data, the current power resource status and power demand can be understood, and operators can formulate corresponding power dispatching or demand response strategies.

[0087] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of the invention.

Claims

1. An enhanced residential load forecasting method based on event response knowledge guidance, characterized in that, The method includes: S1. Obtain historical power sequence data, which includes historical time, power data corresponding to the historical time, target time, and power data corresponding to the target time. S2. Perform state identification on each power sequence data, and obtain the historical time state label corresponding to each historical time and the target time state label corresponding to each target time. S3. Using historical time and the corresponding power data as input, and the target time state label as output, train the first neural network to obtain the power usage event prediction module. The loss function of the power usage event prediction module is the first loss function. S4. Using historical time and corresponding power data as input, and target time and corresponding power data as output, train the second neural network. Adjust the training process of the second neural network using the state probability corresponding to the target time state label output by the power usage event prediction module, the prediction output of the second neural network, and the second loss function calculated from the actual target value to obtain the power prediction model. The adjustment formula is: in, Indicates by The error between the predicted and actual values ​​within the space specified by the logical value. This represents the general regression loss. This represents the final loss of the load prediction module. Indicates the number of dimension variables. Represents the time in the power sequence data. Indicates the prediction time steps. Indicates the first At the time step logits of each dimension variable value, Indicates the first i At the time step Predicted values ​​of each dimension variable, Indicates the first At the time step The actual values ​​of each dimension variable This represents the predicted values ​​for all time steps and variables. This represents the actual values ​​of all time steps and variables. For hyperparameters; S5. Input the predicted power sequence data into the power prediction model, and output the enhanced residential load prediction data through the power prediction model.

2. The method according to claim 1, characterized in that, In step S2, the historical time status label and the target time status label respectively include: ON, OFF, and high performance HP.

3. The method according to claim 1 or 2, characterized in that, Step S2 includes: A state recognition algorithm is used on the power data corresponding to the historical time and the power data corresponding to the target time to determine the historical time state label for each historical time and the target time state label for each target time.

4. The method according to claim 1, characterized in that, Step S4 includes: S41. Determine the state probability based on the target time state label output by the power usage event prediction module; S42. The state probability is used as the weight coefficient in the adjustment formula, and the training process of the second neural network is adjusted by combining the predicted output of the second neural network and the second loss function calculated by the actual target value.

5. The method according to claim 1, characterized in that, In step S4, the second loss function is: in, Indicates by The error between the predicted and actual values ​​within the space specified by the logical value. Represents the actual value of the power sequence data. This represents the predicted value of the power series data. This represents the output of the last layer of the neural network in the power usage event prediction module.

6. The method according to claim 3, characterized in that, The state recognition algorithm includes the k-means clustering algorithm.

7. The method according to claim 1, characterized in that, In step S3, the first neural network includes: an input layer, a convolutional layer, a univariate feature extraction layer, a prediction layer, and an output layer.

8. The method according to claim 1, characterized in that, In step S3, the first loss function is: in, Indicates the number of dimension variables. Indicates the prediction time steps. Indicates in Time of the first The actual state category of each variable Indicates in Time of the first The probability value of each variable being in the actual state category.

Images