Power load prediction method, system and device based on neural network algorithm, medium and computer program product

Abstract

The invention discloses a power load prediction method, system and device based on a neural network algorithm, a medium and a computer program product, solves the problem that an existing power system is low in real-time scheduling and resource allocation precision, and belongs to the field of power system scheduling and management. Comprising the following steps: acquiring target basic data required by power load prediction, preprocessing and dividing into a training set and a test set; associating the Prophet time sequence decomposition model with the CNN-LSTM composite model, and constructing a power load prediction model; training the power load prediction model by using the training set, inputting the test set into the trained power load prediction model, and outputting a power load prediction value; selecting a mean absolute percentage error model, a mean square error model and / or a root-mean-square error model to evaluate the power load predicted value; scheduling and adjusting an allocation strategy of power resources in real time according to an evaluation result; according to the invention, the real-time scheduling and resource allocation precision of the power system is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of power system dispatching and management, and relates to a power load forecasting method, system, device, medium and computer program product based on a neural network algorithm, which is suitable for power load forecasting and optimal dispatching to improve the configuration efficiency of power resources and the stability of the power system. Background Art

[0002] Power load forecasting is an important means for power companies to optimize power dispatch, balance power supply and demand, and reduce costs.

[0003] Most existing power load forecasting methods rely on traditional statistical models and time series analysis. Although they can make predictions to a certain extent, they have low accuracy and reliability when dealing with complex load fluctuations, external environmental changes, and production plan adjustments. Especially in the context of changing power demand, traditional methods often face problems such as large prediction errors and inability to adapt to changes, resulting in inaccurate power dispatch and serious waste of energy resources.

[0004] Existing power load forecasting methods usually use traditional methods based on regression analysis, time series analysis, grey system theory, etc. These methods can handle short-term changes in load, but their accuracy and reliability are greatly reduced when faced with complex and nonlinear changes in power demand. In particular, for scenarios that include a large number of cyclical factors, such as seasonality, holidays, etc., and complex production scheduling, such as industrial enterprise production plans, traditional methods have poor forecasting effects and are difficult to adapt to emergencies or external interference.

[0005] In addition, some prediction models in the existing technology still lack sufficient robustness in environments with missing data, noise or drastic changes, resulting in low prediction accuracy and difficulty in meeting the precise requirements of existing power systems for real-time scheduling and resource allocation. Summary of the invention

[0006] In order to solve the technical problem of low accuracy of real-time scheduling and resource allocation in the existing power system, the present invention provides a power load forecasting method, system, device, medium and computer program product based on a neural network algorithm. By combining the Prophet time series decomposition model with the CNN-LSTM composite model, it can effectively respond to emergencies and external influencing factors, and significantly improve the accuracy of power load forecasting and the stability of the system. In particular, when facing complex power demand scenarios, such as production plan fluctuations, holiday changes, etc., it can capture the dynamic change law of the load in real time and accurately, predict the power load, improve the robustness and adaptability of the prediction, optimize the allocation of power resources, optimize the load forecasting effect in complex power consumption scenarios, and improve the dispatching efficiency and economic benefits of the power system.

[0007] The purpose of the present invention is specifically achieved through the following technical solutions:

[0008] The present invention discloses a method for predicting electric load based on a neural network algorithm, the method comprising:

[0009] Step 1: Obtain target basic data required for power load forecasting from different data sources;

[0010] Step 2: After the target basic data is preprocessed by data cleaning, it is divided into a training set and a test set; according to the task requirements, the corresponding features of the target basic data in the training set and the test set are extracted, and the extracted features are subjected to data standardization to obtain the target data;

[0011] Step 3: Use the output data of the Prophet time series decomposition model as one of the input data of the CNN-LSTM composite model, associate the Prophet time series decomposition model with the CNN-LSTM composite model, and obtain the constructed power load forecasting model; use the target data in the training set as input data to train the power load forecasting model, and obtain the trained power load forecasting model;

[0012] Step 4: Input the target data in the test set into the trained power load forecasting model and output the power load forecast value; select the mean absolute percentage error model, mean square error model and / or root mean square error model to evaluate the power load forecast value and obtain the evaluation result; according to the evaluation result, dispatch and adjust the power resource allocation strategy in real time.

[0013] In step 1, the target basic data includes: real-time power load data, environmental data and production plan data; among which,

[0014] Real-time power load data, which is real-time power load data corresponding to time changes;

[0015] Environmental data, which refers to the environmental data corresponding to external influencing factors;

[0016] The production plan data is the production plan data corresponding to the production plan information.

[0017] In step 2, the preprocessing method for cleaning the target basic data includes:

[0018] Use the 3σ principle or box plot method to identify outliers in the target basic data, and correct or eliminate outliers in the target basic data;

[0019] Use linear interpolation or similar period data to fill in missing values ​​in the target basic data;

[0020] Wavelet transform is used to remove high-frequency noise in the target basic data and perform noise reduction processing.

[0021] In step 2, the method of dividing into training set and test set includes:

[0022] The target basic data after preprocessing operations are divided into a training set and a test set according to chronological order and preset ratio.

[0023] In step 2, the method of extracting corresponding features from the target basic data in the training set and the test set according to the task requirements includes: extracting corresponding time series features, environmental features and / or production features from the target basic data in the training set and the test set according to the task requirements; wherein,

[0024] Time series features are time series features extracted from real-time power load data that change over time, including but not limited to daily cycle features, weekly cycle features and / or holiday features;

[0025] Environmental characteristics are environmental characteristics corresponding to external influencing factors extracted from environmental data, including but not limited to temperature characteristics, humidity characteristics, temperature change trend characteristics and / or humidity change trend characteristics;

[0026] Production characteristics are production characteristics corresponding to production plan information extracted from production plan data, including but not limited to production schedule information characteristics, equipment operation status information characteristics and / or shutdown plan information characteristics.

[0027] In step 2, the method of performing data standardization processing on the extracted features includes:

[0028] The extracted numerical features are normalized using Z-score;

[0029] The extracted categorical features are standardized using One-Hot Encoding.

[0030] In step 3, the Prophet time series decomposition model is:

[0031] y(t)=g(t)+s(t)+h(t)+∈ t

[0032] Where y(t) is the target value, t is the time; g(t) is the trend component, which is used to fit the non-periodic trend change (upward and downward trend) of the time series characteristics; s(t) is the periodic component, which is used to fit the periodic change trend of weeks, months and seasons; h(t) is the holiday component, which is used to indicate the impact of potential jump points on the prediction (such as holidays and emergencies); ∈ tis the error component, which is used to represent the unpredicted random fluctuations (unpredictable part); among them, g(t) and s(t) output by the Prophet time series decomposition model are used as one of the inputs of the CNN-LSTM composite model.

[0033] In step 3, the model architecture of the CNN-LSTM composite model includes a sequentially connected input layer, a CNN layer, and a LSTM layer, where:

[0034] The input layer is used to use the g(t)+s(t) output by the Prophet time series decomposition model and the target data as the input data of the input layer;

[0035] The CNN layer is used to extract the local characteristics of short-term fluctuations in power load in the input data and capture the short-term fluctuations in power load caused by daily fluctuations and production plan adjustments;

[0036] The LSTM layer is used to model the long-term dependencies of local features of short-term fluctuations in power load and capture the dynamic impact of emergencies or environmental changes on power load.

[0037] In step 3, the calculation formula of the CNN layer is:

[0038]

[0039] In the formula, z i is the convolution output feature; k is the total number of convolution kernels, j is the serial number of the convolution kernel; x a is the input sequence, a is the serial number of the input sequence; ω j is the convolution kernel weight; b is the bias term.

[0040] In step 3, the calculation formula of the LSTM layer is:

[0041] f t =σ(W f ·[h t-1 ,x t ]+b f );

[0042] i t =σ(W i ·[h t-1 ,x t ]+b i );

[0043]

[0044]

[0045] o t =σ(W o ·[h t-1 ,xt ]+b o );

[0046] h t =o t tanh(C t );

[0047] In the formula, f t is the output of the forget gate, which is used to control the retention degree of the previous memory, t is the time; σ is the Sigmoid activation function, which compresses the value to (0,1); W f is the forget gate weight matrix, which is used to perform linear transformation on the input and hidden states, f is the forget gate, which determines how much historical information to retain; h t-1 is the hidden state of the previous time step, which is short-term memory; x t is the input feature vector of the current time step; b f is the forget gate bias term;

[0048] i t is the input gate output, which is used to control the degree of adoption of the current input; W i is the input gate weight matrix; b i is the input gate bias term, i is the input gate, which controls the impact of the current input on the memory unit;

[0049] is the candidate memory, which is used to generate the new memory candidate value of the current time step; tanh is the hyperbolic tangent function, which compresses the value to (-1,1); W c is the candidate memory weight matrix, b c is the candidate memory bias item, c is the candidate memory, indicating the new memory candidate value, and is ready to update the memory unit state at the current moment;

[0050] C t is the memory unit state at the current time step, which is long-term memory; C t-1 is the state of the memory unit at the previous time step, which is short-term memory;

[0051] o t is the output gate output, which is used to control the output degree of the hidden state; W o is the output gate weight matrix; b o is the output gate bias term, o is the output gate, which determines the hidden state content output at the current moment;

[0052] h t is the hidden state of the current time step, which is short-term memory.

[0053] In step 3, the target data in the training set is used as input data to train the power load forecasting model, and the method for obtaining the trained power load forecasting model includes:

[0054] S1, input the target data of the training set into the Prophet time series decomposition model, and output the trend component g(t) and periodic component s(t) of the power load data forecast;

[0055] S2, the trend component g(t) and periodic component s(t) output by the Prophet time series decomposition model and the target data of the training set are simultaneously used as the input data of the CNN-LSTM composite model; the short-term fluctuation local characteristics of the power load in the input data are extracted through the CNN layer of the CNN-LSTM composite model; the long-term dependency of the short-term fluctuation local characteristics of the power load is modeled through the LSTM layer of the CNN-LSTM composite model; the backpropagation algorithm is used to calculate the gradient; the Adam optimizer is used to update the model parameters, and the power load prediction value is output to obtain the trained power load prediction model.

[0056] In step 4, the method for evaluating the output power load forecast value by using the mean absolute percentage error model includes:

[0057]

[0058] In the formula, MAPE is the mean absolute percentage error, n is the total number of power loads, and e is the serial number of the power load. is the predicted value of power load, y e is the actual value of the power load.

[0059] In step 4, the method for evaluating the output power load forecast value through the mean square error model includes:

[0060]

[0061] In the formula, MSE is the mean square error, n is the total number of power loads, and e is the serial number of the power load. is the predicted value of power load, y e is the actual value of the power load.

[0062] In step 4, the method for evaluating the output power load forecast value through the root mean square error model includes:

[0063]

[0064] In the formula, RMSE is the root mean square error, n is the total number of power loads, e is the serial number of power loads, is the power load forecast value, y e is the actual value of the power load.

[0065] The present invention also provides a power load forecasting system based on a neural network algorithm, comprising:

[0066] Data acquisition module, used to obtain target basic data required for power load forecasting from different data sources;

[0067] The data preprocessing and feature extraction module is used to divide the target basic data into training sets and test sets after data cleaning preprocessing operations; extract corresponding features from the target basic data in the training set and test set according to task requirements, and obtain the target data after data standardization processing of the extracted features;

[0068] The power load forecasting model construction module is used to use the output data of the Prophet time series decomposition model as one of the input data of the CNN-LSTM composite model, associate the Prophet time series decomposition model with the CNN-LSTM composite model, and obtain the constructed power load forecasting model; use the target data in the training set as input data to train the power load forecasting model, and obtain the trained power load forecasting model;

[0069] The scheduling decision module is used to input the target data in the test set into the trained power load forecasting model and output the power load forecast value; select the mean absolute percentage error model, the mean square error model and / or the root mean square error model to evaluate the power load forecast value to obtain the evaluation result; and according to the evaluation result, real-time scheduling and adjustment of the power resource allocation strategy.

[0070] The present invention also provides an electric power load prediction device based on a neural network algorithm, comprising a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method.

[0071] The present invention also provides a computer-readable storage medium for power load forecasting based on a neural network algorithm, on which a computer program / instruction is stored, and the steps of the method are implemented when the computer program / instruction is executed by a processor.

[0072] The present invention also provides a computer program product for power load forecasting based on a neural network algorithm, comprising a computer program / instruction, which implements the steps of the method when executed by a processor.

[0073] The beneficial effects of the present invention are:

[0074] 1. Improve prediction accuracy

[0075] By combining the Prophet time series decomposition model with the CNN-LSTM composite model, the constructed power load forecasting model can accurately capture the long-term trend, seasonal fluctuations and complex short-term fluctuations of power load, especially when facing the influence of external factors (such as holidays and changes in production plans), it can still maintain a high prediction accuracy.

[0076] 2. Strong adaptability

[0077] The electric power load forecasting model constructed by the present invention has strong adaptive capability, can adjust itself according to the changes of real-time load data, optimize the forecasting performance, and ensure the long-term stability and high accuracy of the electric power load forecasting.

[0078] 3. Enhanced robustness

[0079] When faced with data loss, noise interference or emergencies, the power load forecasting model constructed by the present invention can still maintain stable forecasting performance and demonstrate strong robustness.

[0080] 4. Applicable in multiple scenarios

[0081] The present invention is applicable to various power load forecasting needs such as residential, commercial and industrial, and can be flexibly configured according to different application scenarios to provide more accurate load forecasting results.

[0082] 5. Optimize power dispatch

[0083] Through the evaluation results of the power load forecast value, the present invention can provide a scientific and real-time load demand forecast for the power dispatching system, ensure the rational allocation of power resources, reduce energy waste, and improve the real-time dispatching and resource allocation accuracy of the existing power system as well as the overall efficiency of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] The present invention is further described in detail below based on the accompanying drawings and embodiments.

[0085] Figure 1 It is a schematic diagram of the experimental results of the normal time period provided in Example 1 of the present invention.

[0086] Figure 2 It is a schematic diagram of the experimental results of a special period provided in Example 1 of the present invention. DETAILED DESCRIPTION

[0087] The embodiment of the present invention provides a method for predicting power load based on a neural network algorithm, the method comprising:

[0088] Step 1: Obtain target basic data required for power load forecasting from different data sources;

[0089] Step 2: After the target basic data is preprocessed by data cleaning, it is divided into a training set and a test set; according to the task requirements, the corresponding features of the target basic data in the training set and the test set are extracted, and the extracted features are subjected to data standardization to obtain the target data;

[0090] Step 3: Use the output data of the Prophet time series decomposition model as one of the input data of the CNN-LSTM composite model, associate the Prophet time series decomposition model with the CNN-LSTM composite model, and obtain the constructed power load forecasting model; use the target data in the training set as input data to train the power load forecasting model, and obtain the trained power load forecasting model;

[0091] Step 4: Input the target data in the test set into the trained power load forecasting model and output the power load forecast value; select the mean absolute percentage error model, mean square error model and / or root mean square error model to evaluate the power load forecast value and obtain the evaluation result; according to the evaluation result, dispatch and adjust the power resource allocation strategy in real time.

[0092] By coordinating with the power dispatching system, the smooth operation of the power system is ensured to avoid resource waste or overload.

[0093] In step 1, the target basic data includes: real-time power load data, environmental data and production plan data; among which,

[0094] Real-time power load data refers to real-time power load data corresponding to time changes; for example, daily and hourly power load data are obtained through a power monitoring system.

[0095] Environmental data refers to environmental data corresponding to external influencing factors; for example, external influencing factors such as temperature, humidity, weather conditions, and holidays are collected from the environmental monitoring system.

[0096] The production plan data is the production plan data corresponding to the production plan information, such as the production schedule, equipment operation status and shutdown plan of the enterprise obtained from the production scheduling system.

[0097] In step 2, the preprocessing method for cleaning the target basic data includes:

[0098] Use the 3σ principle or box plot method to identify outliers in the target basic data, and correct or eliminate outliers in the target basic data;

[0099] Use linear interpolation or similar period data to fill in missing values ​​in the target basic data;

[0100] Wavelet transform is used to remove high-frequency noise in the target basic data and perform noise reduction processing.

[0101] In step 2, the method of dividing into training set and test set includes:

[0102] The target basic data after preprocessing operations are divided into training set and test set according to the time sequence and preset ratio. For example, the data set is divided into 80% training set and 20% test set according to the time sequence, ensuring that the test set does not participate in the training.

[0103] In step 2, the method of extracting corresponding features from the target basic data in the training set and the test set according to the task requirements includes: extracting corresponding time series features, environmental features and / or production features from the target basic data in the training set and the test set according to the task requirements; wherein,

[0104] Time series features are time series features extracted from real-time power load data that change over time, including but not limited to daily cycle features, weekly cycle features and / or holiday features;

[0105] Environmental characteristics are environmental characteristics corresponding to external influencing factors extracted from environmental data, including but not limited to temperature characteristics, humidity characteristics, temperature change trend characteristics and / or humidity change trend characteristics;

[0106] Production characteristics are production characteristics corresponding to production plan information extracted from production plan data, including but not limited to production schedule information characteristics, equipment operation status information characteristics and / or shutdown plan information characteristics.

[0107] In step 2, the method of performing data standardization processing on the extracted features includes:

[0108] The extracted numerical features are normalized using Z-score to ensure that the numerical feature range is consistent;

[0109] The extracted categorical features, such as holiday labels, are standardized using One-Hot Encoding.

[0110] In step 3, the Prophet time series decomposition model is used to model the long-term trend and periodic fluctuation of power load. The Prophet time series decomposition model is:

[0111] y(t)=g(t)+s(t)+h(t)+∈ t

[0112] Where y(t) is the target value, t is the time; g(t) is the trend component, which is used to fit the non-periodic trend change (upward and downward trend) of the time series characteristics; s(t) is the periodic component, which is used to fit the periodic change trend of weeks, months and seasons; h(t) is the holiday component, which is used to indicate the impact of potential jump points on the prediction (such as holidays and emergencies); ∈ t is the error component, which is used to represent the unpredicted random fluctuations (unpredictable part); among them, g(t) and s(t) output by the Prophet time series decomposition model are used as one of the inputs of the CNN-LSTM composite model.

[0113] In step 3, the CNN-LSTM composite model further explores the short-term fluctuation characteristics and long-term dependencies of power load. The model architecture includes a sequentially connected input layer, a CNN layer, and a LSTM layer.

[0114] The input layer is used to use the g(t)+s(t) output by the Prophet time series decomposition model and the target data as the input data of the input layer;

[0115] The CNN layer is used to extract the local characteristics of short-term fluctuations in power load in the input data and capture the short-term fluctuations in power load caused by daily fluctuations and production plan adjustments;

[0116] The LSTM layer is used to model the long-term dependencies of local features of short-term fluctuations in power load and capture the dynamic impact of emergencies or environmental changes on power load.

[0117] In step 3, the calculation formula of the CNN layer is:

[0118]

[0119] In the formula, z i is the convolution output feature; k is the total number of convolution kernels, j is the serial number of the convolution kernel; x a is the input sequence, a is the serial number of the input sequence; ω j is the convolution kernel weight; b is the bias term.

[0120] In step 3, the calculation formula of the LSTM layer is:

[0121] f t =σ(W f ·[h t-1 ,x t ]+b f );

[0122] i t =σ(W i ·[h t-1 ,x t ]+bi );

[0123]

[0124] o t =σ(W o ·[h t-1 ,x t ]+b o );

[0125] h t =o t tanh(C t );

[0126] In the formula, f t is the output of the forget gate, which is used to control the retention degree of the previous memory, t is the time; σ is the Sigmoid activation function, which compresses the value to (0,1); W f is the forget gate weight matrix, which is used to perform linear transformation on the input and hidden states, f is the forget gate, which determines how much historical information to retain; h t-1 is the hidden state of the previous time step, which is short-term memory; x t is the input feature vector of the current time step; b f is the forget gate bias term;

[0127] i t is the input gate output, which is used to control the degree of adoption of the current input; W i is the input gate weight matrix; b i is the input gate bias term, i is the input gate, which controls the impact of the current input on the memory unit;

[0128] is the candidate memory, which is used to generate the new memory candidate value of the current time step; tanh is the hyperbolic tangent function, which compresses the value to (-1,1); W c is the candidate memory weight matrix, b c is the candidate memory bias item, c is the candidate memory, indicating the new memory candidate value, and is ready to update the memory unit state at the current moment;

[0129] C t is the memory unit state at the current time step, which is long-term memory; C t-1 is the state of the memory unit at the previous time step, which is short-term memory;

[0130] o t is the output gate output, which is used to control the output degree of the hidden state; W o is the output gate weight matrix; b o is the output gate bias term, o is the output gate, which determines the hidden state content output at the current moment;

[0131] h t is the hidden state of the current time step, which is short-term memory.

[0132] In step 3, the target data in the training set is used as input data to train the power load forecasting model, and the method for obtaining the trained power load forecasting model includes:

[0133] S1, input the target data of the training set into the Prophet time series decomposition model, and output the trend component g(t) and periodic component s(t) of the power load data forecast;

[0134] S2, the trend component g(t) and periodic component s(t) output by the Prophet time series decomposition model and the target data of the training set are simultaneously used as the input data of the CNN-LSTM composite model; the short-term fluctuation local characteristics of the power load in the input data are extracted through the CNN layer of the CNN-LSTM composite model; the long-term dependency of the short-term fluctuation local characteristics of the power load is modeled through the LSTM layer of the CNN-LSTM composite model; the backpropagation algorithm is used to calculate the gradient; the Adam optimizer is used to update the model parameters and output the power load prediction value, including the comprehensive prediction of short-term fluctuations, periodic changes and long-term trends; and a trained power load prediction model is obtained.

[0135] In step 4, the method for evaluating the output power load forecast value by using the mean absolute percentage error model includes:

[0136]

[0137] In the formula, MAPE is the mean absolute percentage error, n is the total number of power loads, and e is the serial number of the power load. is the power load forecast value, y e is the actual value of the power load.

[0138] In step 4, the method for evaluating the output power load forecast value through the mean square error model includes:

[0139]

[0140] In the formula, MSE is the mean square error, n is the total number of power loads, and e is the serial number of the power load. is the predicted value of power load, y e is the actual value of the power load.

[0141] In step 4, the method for evaluating the output power load forecast value through the root mean square error model includes:

[0142]

[0143] In the formula, RMSE is the root mean square error, n is the total number of power loads, e is the serial number of power loads, is the power load forecast value, y e is the actual value of the power load.

[0144] The combination of the Prophet time series decomposition model and the CNN-LSTM composite model in the present invention adopts staged modeling, making full use of the advantages of the two models and improving the prediction accuracy and robustness through data flow integration. The specific combination method is as follows:

[0145] Phase 1: Prophet time series decomposition model extracts macro features

[0146] (1) Function:

[0147] The Prophet time series decomposition model is mainly responsible for capturing the long-term trend and periodic fluctuations of load data and reducing the complexity of the original load data.

[0148] By performing trend decomposition on time series data, long-term trend components and cyclical components are output, providing clearer macro guidance for subsequent short-term feature modeling.

[0149] (2) Output:

[0150] The g(t) and s(t) output by the Prophet time series decomposition model are used as one of the input features of the CNN-LSTM composite model and combined with the original load data and other auxiliary features (such as environmental data).

[0151] Phase 2: CNN-LSTM composite model mining local features and dependencies

[0152] (1) Function:

[0153] Based on the long-term features provided by the Prophet time series decomposition model, the CNN-LSTM composite model further learns the short-term fluctuation patterns and time series dependencies of load data.

[0154] The CNN layer is used to extract local features and capture the short-term impact of daily fluctuations and production plan adjustments on load.

[0155] The LSTM layer is responsible for modeling the long-term dependencies of the time series and capturing the dynamic impact of sudden events or environmental changes on the load.

[0156] (2) Combination points:

[0157] The g(t) and s(t) extracted by the Prophet time series decomposition model are combined with the original load data and other features as the input of the CNN-LSTM composite model.

[0158] After the CNN layer extracts local patterns, the LSTM layer further models the global time series dependencies.

[0159] Phase 3: Final prediction output

[0160] (1) Integration method:

[0161] The CNN-LSTM composite model outputs the load forecast value y^(t) through the trend information (g(t), s(t)) provided by the Prophet time series decomposition model and the short-term fluctuation characteristics learned by itself.

[0162] The comprehensive forecast results include short-term fluctuations, cyclical changes and long-term trend information.

[0163] (2) Advantages:

[0164] This staged modeling approach enables the Prophet time series decomposition model and the CNN-LSTM composite model to complement each other:

[0165] a. The Prophet time series decomposition model provides a macro perspective and enhances the trend and periodicity modeling capabilities.

[0166] b. The CNN-LSTM composite model optimizes short-term dynamic prediction and enhances the model’s adaptability to complex scenarios.

[0167] The embodiment of the present invention further provides a power load forecasting system based on a neural network algorithm, comprising:

[0168] Data acquisition module, used to obtain target basic data required for power load forecasting from different data sources;

[0169] The data preprocessing and feature extraction module is used to divide the target basic data into training sets and test sets after data cleaning preprocessing operations; extract corresponding features from the target basic data in the training set and test set according to task requirements, and obtain the target data after data standardization processing of the extracted features;

[0170] The power load forecasting model construction module is used to use the output data of the Prophet time series decomposition model as one of the input data of the CNN-LSTM composite model, associate the Prophet time series decomposition model with the CNN-LSTM composite model, and obtain the constructed power load forecasting model; use the target data in the training set as input data to train the power load forecasting model, and obtain the trained power load forecasting model;

[0171] The scheduling decision module is used to input the target data in the test set into the trained power load forecasting model and output the power load forecast value; select the mean absolute percentage error model, the mean square error model and / or the root mean square error model to evaluate the power load forecast value to obtain the evaluation result; and according to the evaluation result, real-time scheduling and adjustment of the power resource allocation strategy.

[0172] An embodiment of the present invention also provides an electric load prediction device based on a neural network algorithm, comprising a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of an electric load prediction method based on a neural network algorithm.

[0173] An embodiment of the present invention also provides a computer-readable storage medium for power load prediction based on a neural network algorithm, on which a computer program / instruction is stored. When the computer program / instruction is executed by a processor, the steps of a power load prediction method based on a neural network algorithm are implemented.

[0174] An embodiment of the present invention also provides a computer program product for power load prediction based on a neural network algorithm, including a computer program / instructions, which, when executed by a processor, implements the steps of a power load prediction method based on a neural network algorithm.

[0175] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices, media and computer program products described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0176] In order to further explain the technical solution provided by the present invention, a specific example is now provided to explain the technical solution of the present invention in detail:

[0177] Example 1: Shenzhen Virtual Power Plant Project

[0178] In order to verify the effectiveness of the present invention, the historical electricity consumption data of an industrial enterprise is used for experiments. The enterprise performs load forecasting on different working days, holidays, and production plan fluctuations (such as changes in production tasks). Figure 1 to Figure 2 As shown, the experimental results show that, through the power load forecasting device of the present invention, the forecast error is about 5% in normal time periods, and the forecast error is less than 10% in special time periods (such as holidays and production scheduling adjustments), which is significantly better than the traditional statistical model.

[0179] Implementation steps:

[0180] 1. Obtain historical power consumption data, production plans and environmental data from the power monitoring system and production scheduling system;

[0181] 2. Preprocess the data to extract features such as periodicity and time-frequency;

[0182] 3. Use the training set to train the Prophet+CNN-LSTM model;

[0183] 4. Periodically output the power load forecast results through the load forecast module;

[0184] 5. Adjust the power dispatch strategy according to the forecast results to ensure the rational allocation of power resources.

[0185] From this example we can conclude that:

[0186] 1. The power load forecasting model built based on Prophet+CNN-LSTM, combined with time series decomposition algorithms and two deep learning technologies, can accurately predict power load.

[0187] 2. Strong adaptability and robustness, able to handle complex load fluctuations and changes in external factors (such as production plan adjustments, holiday impacts, etc.).

[0188] 3. Realize the optimal dispatch of power resources, reduce energy waste and improve the efficiency of the power system.

[0189] The beneficial effects of the embodiments of the present invention are:

[0190] 1. Improve prediction accuracy

[0191] By combining the Prophet time series decomposition model with the CNN-LSTM composite model, the constructed power load forecasting model can accurately capture the long-term trend, seasonal fluctuations and complex short-term fluctuations of power load, especially when facing the influence of external factors (such as holidays and changes in production plans), it can still maintain a high prediction accuracy.

[0192] 2. Strong adaptability

[0193] The electric power load forecasting model constructed by the present invention has strong adaptive capability, can adjust itself according to the changes of real-time load data, optimize the forecasting performance, and ensure the long-term stability and high accuracy of the electric power load forecasting.

[0194] 3. Enhanced robustness

[0195] When faced with data loss, noise interference or emergencies, the power load forecasting model constructed by the present invention can still maintain stable forecasting performance and demonstrate strong robustness.

[0196] 4. Applicable in multiple scenarios

[0197] The present invention is applicable to various power load forecasting needs such as residential, commercial and industrial, and can be flexibly configured according to different application scenarios to provide more accurate load forecasting results.

[0198] 5. Optimize power dispatch

[0199] Through the evaluation results of the power load forecast value, the present invention can provide a scientific and real-time load demand forecast for the power dispatching system, ensure the rational allocation of power resources, reduce energy waste, and improve the real-time dispatching and resource allocation accuracy of the existing power system as well as the overall efficiency of the power system.

[0200] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A method for predicting power load based on a neural network algorithm, characterized in that: The method includes: Step 1: Obtain target basic data required for power load forecasting from different data sources; Step 2: After the target basic data is preprocessed by data cleaning, it is divided into a training set and a test set; according to the task requirements, the corresponding features of the target basic data in the training set and the test set are extracted, and the extracted features are subjected to data standardization to obtain the target data; Step 3: Use the output data of the Prophet time series decomposition model as one of the input data of the CNN-LSTM composite model, associate the Prophet time series decomposition model with the CNN-LSTM composite model, and obtain the constructed power load forecasting model; use the target data in the training set as input data to train the power load forecasting model, and obtain the trained power load forecasting model; Step 4: Input the target data in the test set into the trained power load forecasting model and output the power load forecast value; select the mean absolute percentage error model, mean square error model and / or root mean square error model to evaluate the power load forecast value and obtain the evaluation result; according to the evaluation result, dispatch and adjust the power resource allocation strategy in real time.

2. The method according to claim 1, characterized in that In step 1, the target basic data includes: real-time power load data, environmental data and production plan data; among which, Real-time power load data, which is real-time power load data corresponding to time changes; Environmental data, which refers to the environmental data corresponding to external influencing factors; The production plan data is the production plan data corresponding to the production plan information.

3. The method according to claim 2, characterized in that In step 2, the preprocessing method for cleaning the target basic data includes: Use the 3σ principle or box plot method to identify outliers in the target basic data, and correct or eliminate outliers in the target basic data; Use linear interpolation or similar period data to fill in missing values ​​in the target basic data; Wavelet transform is used to remove high-frequency noise in the target basic data and perform noise reduction processing.

4. The method according to claim 3, characterized in that In step 2, the method of dividing into training set and test set includes: The target basic data after preprocessing operations are divided into a training set and a test set according to chronological order and preset ratio.

5. The method according to claim 4, characterized in that In step 2, the method of extracting corresponding features from the target basic data in the training set and the test set according to the task requirements includes: extracting corresponding time series features, environmental features and / or production features from the target basic data in the training set and the test set according to the task requirements; wherein, Time series features are time series features extracted from real-time power load data that change over time, including but not limited to daily cycle features, weekly cycle features and / or holiday features; Environmental characteristics are environmental characteristics corresponding to external influencing factors extracted from environmental data, including but not limited to temperature characteristics, humidity characteristics, temperature change trend characteristics and / or humidity change trend characteristics; Production characteristics are production characteristics corresponding to production plan information extracted from production plan data, including but not limited to production schedule information characteristics, equipment operation status information characteristics and / or shutdown plan information characteristics.

6. The method according to claim 5, characterized in that In step 2, the method of performing data standardization processing on the extracted features includes: The extracted numerical features are normalized using Z-score; The extracted categorical features are standardized using one-hot encoding.

7. The method according to claim 1, characterized in that In step 3, the Prophet time series decomposition model is: y(t)=g(t)+s(t)+h(t)+∈ t In the formula, y(t) is the target value; t is time; g(t) is the trend component, which is used to fit the non-periodic trend change of time series characteristics; s(t) is the periodic component, which is used to fit the periodic change trend of weeks, months and seasons; h(t) is the holiday component, which is used to indicate the impact of potential jump points on prediction; ∈ t is the error component, which is used to represent the unpredicted random fluctuations; among them, g(t) and s(t) output by the Prophet time series decomposition model are used as one of the inputs of the CNN-LSTM composite model.

8. The method according to claim 1 or 7, characterized in that In step 3, the model architecture of the CNN-LSTM composite model includes a sequentially connected input layer, a CNN layer, and a LSTM layer, where: The input layer is used to use the g(t)+s(t) output by the Prophet time series decomposition model and the target data as the input data of the input layer; The CNN layer is used to extract the local characteristics of short-term fluctuations in power load in the input data and capture the short-term fluctuations in power load caused by daily fluctuations and production plan adjustments; The LSTM layer is used to model the long-term dependencies of local features of short-term fluctuations in power load and capture the dynamic impact of emergencies or environmental changes on power load.

9. The method according to claim 8, characterized in that In step 3, the calculation formula of the CNN layer is: In the formula, z i is the convolution output feature; k is the total number of convolution kernels, j is the serial number of the convolution kernel; x a is the input sequence, a is the serial number of the input sequence; ω j is the convolution kernel weight; b is the bias term.

10. The method according to claim 8, characterized in that In step 3, the calculation formula of the LSTM layer is: f t =σ(W f ·[h t-1 ,x t ]+b f ); i t =σ(W i ·[h t-1 ,x t ]+b i ); the t =σ(W o ·[h t-1 ,x t ]+b o ); h t =o t ·tanh(C t ); In the formula, f t is the output of the forget gate, which is used to control the retention degree of the previous memory, t is the time; σ is the Sigmoid activation function, which compresses the value to (0,1); W f is the forget gate weight matrix, which is used to perform linear transformation on the input and hidden states, f is the forget gate, which determines how much historical information to retain; h t-1 is the hidden state of the previous time step, which is short-term memory; x t is the input feature vector of the current time step; b f is the forget gate bias term; i t is the input gate output, which is used to control the degree of adoption of the current input; W i is the input gate weight matrix; b i is the input gate bias term, i is the input gate, which controls the impact of the current input on the memory unit; is the candidate memory, which is used to generate the new memory candidate value of the current time step; tanh is the hyperbolic tangent function, which compresses the value to (-1,1); W c is the candidate memory weight matrix, b c is the candidate memory bias item, c is the candidate memory, indicating the new memory candidate value, and is ready to update the memory unit state at the current moment; C t is the memory unit state at the current time step, which is long-term memory; C t-1 is the state of the memory unit at the previous time step, which is short-term memory; o t is the output gate output, which is used to control the output degree of the hidden state; W o is the output gate weight matrix; b o is the output gate bias term, o is the output gate, which determines the hidden state content output at the current moment; h t is the hidden state of the current time step, which is short-term memory.

11. The method according to claim 10, characterized in that In step 3, the target data in the training set is used as input data to train the power load forecasting model, and the method for obtaining the trained power load forecasting model includes: S1, input the target data of the training set into the Prophet time series decomposition model, and output the trend component g(t) and periodic component s(t) of the power load data forecast; S2, the trend component g(t) and periodic component s(t) output by the Prophet time series decomposition model and the target data of the training set are simultaneously used as the input data of the CNN-LSTM composite model; the short-term fluctuation local characteristics of the power load in the input data are extracted through the CNN layer of the CNN-LSTM composite model; the long-term dependency of the short-term fluctuation local characteristics of the power load is modeled through the LSTM layer of the CNN-LSTM composite model; the back propagation algorithm is used to calculate the gradient; the Adam optimizer is used to update the model parameters, and the power load prediction value is output to obtain the trained power load prediction model.

12. The method according to claim 1 or 11, characterized in that In step 4, the method for evaluating the output power load forecast value by using the mean absolute percentage error model includes: In the formula, MAPE is the mean absolute percentage error, n is the total number of power loads, and e is the serial number of the power load. is the predicted value of power load, y e is the actual value of the power load.

13. The method according to claim 1 or 11, characterized in that: In step 4, the method for evaluating the output power load forecast value through the mean square error model includes: In the formula, MSE is the mean square error, n is the total number of power loads, and e is the serial number of the power load. is the predicted value of power load, y e is the actual value of the power load.

14. The method according to claim 1 or 11, characterized in that: In step 4, the method for evaluating the output power load forecast value through the root mean square error model includes: In the formula, RMSE is the root mean square error, n is the total number of power loads, e is the serial number of power loads, is the predicted value of power load, y e is the actual value of the power load.

15. A power load forecasting system based on a neural network algorithm, characterized in that: include: Data acquisition module, used to obtain target basic data required for power load forecasting from different data sources; The data preprocessing and feature extraction module is used to divide the target basic data into training sets and test sets after data cleaning preprocessing operations; extract corresponding features from the target basic data in the training set and test set according to task requirements, and obtain the target data after data standardization processing of the extracted features; The power load forecasting model construction module is used to use the output data of the Prophet time series decomposition model as one of the input data of the CNN-LSTM composite model, associate the Prophet time series decomposition model with the CNN-LSTM composite model, and obtain the constructed power load forecasting model; use the target data in the training set as input data to train the power load forecasting model, and obtain the trained power load forecasting model; The scheduling decision module is used to input the target data in the test set into the trained power load forecasting model and output the power load forecast value; select the mean absolute percentage error model, the mean square error model and / or the root mean square error model to evaluate the power load forecast value to obtain the evaluation result; and according to the evaluation result, real-time scheduling and adjustment of the power resource allocation strategy.

16. A power load forecasting device based on a neural network algorithm, comprising a memory, a processor and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1-14.

17. A computer-readable storage medium for power load forecasting based on a neural network algorithm, having a computer program / instruction stored thereon, characterized in that: When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 14 are implemented.

18. A computer program product for power load forecasting based on a neural network algorithm, comprising a computer program / instructions, characterized in that: When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 14 are implemented.

Images