An outwards-oriented urban power load forecasting method

Abstract

A kind of export-oriented city electric power load forecasting method considering foreign trade prosperity conduction hysteresis, through the cleaning and alignment of multi-source heterogeneous data, time lag analysis based on maximum information coefficient method and through double-channel neural network model prediction, output final future time electric power load prediction value;The double-channel neural network model includes time series fluctuation channel and foreign trade fluctuation channel, mainly rely on time series fluctuation channel when meteorological mutation, and use foreign trade fluctuation channel to adjust load benchmark when external impact is serious.Can adapt to the change of international trade environment, and accurately reflect its hysteresis effect, improve the accuracy of electric power load forecasting result, to help guarantee the safe and stable operation of power system and reliable power supply.

Description

Technical Field

[0001] This invention relates to the field of power load forecasting technology, and in particular to a method for forecasting power load in export-oriented cities that takes into account the lag in the transmission of foreign trade prosperity. Background Technology

[0002] With the accelerated implementation of energy transition strategies and the deepening construction of new power systems characterized by high-proportion renewable energy integration, the strong randomness and volatility of power output from renewable energy sources such as wind and solar have exacerbated the uncertainty of net grid load. Simultaneously, the rapid development of new business models such as grid-side energy storage, electric vehicles, and charging piles, coupled with the frequent occurrence of unpredictable public events and natural disasters, further increases the complexity of load forecasting. In particular, for export-oriented cities, electricity consumption is not only affected by meteorological conditions and seasonal patterns but is also closely related to changes in the international macroeconomic environment. Currently, global geopolitical tensions and frequent changes in tariff policies are impacting industrial production and exports, leading to significant fluctuations in regional grid load. In summary, load forecasting faces unprecedented challenges, necessitating the development of a high-precision power load forecasting method.

[0003] Existing technologies mainly rely on GDP, population size, and temperature for load forecasting, without taking external shocks into account. They also ignore the "time lag" between the signing of foreign trade orders and the electricity consumption for production. This can easily lead to a significant decrease in load forecasting accuracy when faced with a sudden increase or decrease in foreign trade orders, making it impossible to meet the needs of refined power grid dispatching. Summary of the Invention

[0004] In view of the above, in order to solve the problems existing in the above-mentioned technologies, this invention proposes an export-oriented urban power load forecasting method that takes into account the lag in the transmission of foreign trade prosperity. This method can adapt to changes in the international trade environment and accurately reflect its lag effects, thereby improving the accuracy of power load forecasting results and helping to ensure the safe and stable operation of the power system and the reliable supply of electricity.

[0005] The present invention proposes a method for forecasting electricity load in export-oriented cities that considers the lag in the transmission of foreign trade prosperity, specifically including the following contents.

[0006] First, the multi-source heterogeneous data is cleaned and aligned, including the following steps:

[0007] (1) Collect raw data, including: power load data Meteorological data Foreign trade data ;

[0008] (2) Outlier detection: For power load, box plots are used to remove outliers; for meteorological and foreign trade data, extreme fluctuations are retained to capture their impact on power load.

[0009] (3) Multi-scale alignment of foreign trade data: Select daily frequency as the basic frequency for modeling, down-frequency analysis of time lag, and up-frequency construction of features;

[0010] (4) Normalization process: For any original time series The data is normalized.

[0011] Secondly, a time delay analysis based on the maximum information coefficient method is performed, including the following steps:

[0012] (1) Set the maximum conduction lag time K;

[0013] (2) Calculate the maximum information coefficient (MIC). For each input index, calculate the MIC under different transmission lag days. Below, its maximum information coefficient MIC with daily load;

[0014] (3) Determine the optimal conduction lag period That is, the number of days corresponding to the maximum MIC of daily load under different transmission lag days, i.e. The k value corresponding to the maximum value in the middle;

[0015] (4) Feature reconstruction: When inputting data into the prediction model, foreign trade data is not input for the current day, but rather... Data from the past day;

[0016] Finally, the final future power load forecast is output through a dual-channel neural network model. The dual-channel neural network model includes a time-series fluctuation channel and a foreign trade fluctuation channel. It mainly relies on the time-series fluctuation channel during sudden weather changes, while the foreign trade fluctuation channel is used to adjust the load benchmark when external shocks are severe.

[0017] The power load data , representing the daily maximum load of the region on day i, where i = 1, 2, ..., T; the meteorological data These refer to the highest temperature, lowest temperature, average temperature, atmospheric pressure, relative humidity, average wind speed, rainfall, and snowfall in the region on day i, respectively; the foreign trade data... These represent the Baltic Dry Index, the RMB exchange rate of major exporting countries in the region, the China Export Container Freight Index, the China Manufacturing New Export Orders Index, and the regional foreign trade cargo throughput, respectively.

[0018] The frequency reduction analysis time lag is that the input indicators directly use the release frequency to perform statistical calculations on the load; for weekly release indicators, the weekly maximum load is calculated when analyzing the time lag; for monthly release indicators (PMI, PT), the monthly maximum load is calculated by accumulating the daily maximum load; the frequency increase construction feature is that all indicators are converted into daily frequency data.

[0019] To convert all indicators to daily frequency data, the procedure is as follows:

[0020] Daily frequency data, including the Baltic Dry Index and the RMB exchange rate, remain unchanged and can be used directly.

[0021] Weekly data, including the China Export Container Freight Index, uses a forward-filling method;

[0022] Monthly frequency data, including the China Manufacturing New Export Orders Index, uses the forward-filling method;

[0023] Cumulative data, including regional foreign trade cargo throughput, is averaged daily and filled forward. The monthly foreign trade cargo throughput is divided by the number of days in the month to obtain the daily foreign trade cargo throughput for that month, thus eliminating dimensional fluctuations caused by differences in the number of days in the month. When the monthly foreign trade cargo throughput is unknown, the data for that month is not used; instead, the average daily foreign trade cargo throughput of the most recent known month is mapped.

[0024] The method for converting the regional foreign trade cargo throughput indicator into daily frequency data also includes:

[0025] Introducing a daily vessel arrival count index that is directly proportional to the volume of foreign trade cargo, the monthly foreign trade cargo throughput is constructed into a daily frequency series through weighted decomposition.

[0026]

[0027] in, It is the foreign trade cargo throughput on day i. The volume of foreign trade goods in this month, This represents the number of ships arriving on day i.

[0028] The normalization method is as follows:

[0029]

[0030] in, This represents the value on day i after normalization of the original sequence. This represents the value of the original sequence on day i. Represents the minimum value in the sequence. This represents the maximum value in the sequence. The final output is a set of dimensionless, multidimensional time series vectors.

[0031] The dual-channel neural network model includes:

[0032] (1) Channel A: Temporal coding module based on CNN-BiLSTM; constructing time series tensors It includes historical electricity load data for T days, meteorological data, and date type features; the construction method involves, firstly, processing the input data through a convolutional neural network. Local feature extraction is performed to learn short-term local fluctuations in load and meteorological sequences, resulting in a convolutional output. Then, output the convolution. The data is fed into a bidirectional long short-term memory network, which utilizes its forward and backward LSTM layers to acquire sequence context associations. This network learns the daily, weekly, and yearly periodic dependencies between load and weather, ultimately outputting time-series features. ;

[0033] (2) Channel B: Foreign trade coding module based on TCN; constructing a foreign trade indicator tensor aligned by time delay. It includes the Baltic Dry Index, the RMB exchange rate of major exporting countries in the region, the China Export Container Freight Index, the China Manufacturing New Export Orders Index, and the regional foreign trade cargo throughput; the construction method involves using a temporal convolutional network to process the input. This study utilizes causal convolution and dilated convolution to extract the long-term evolution trend of foreign trade prosperity and outputs foreign trade characteristics. ;

[0034] (3) Cross-modal feature fusion module based on attention mechanism; cross-modal feature fusion integrates time series and foreign trade features, the time series features and foreign trade features are respectively and To match the tensor sizes of the two modal features, a query vector is obtained through linear mapping. Key vector Value vector The specific process of implementing the cross-modal attention mechanism is as follows:

[0035]

[0036] In the formula, The final feature vector after fusion. , These are the original feature preservation coefficients and cross-modal correction coefficients, respectively, which are automatically updated during model training. Softmax() is the probabilistic function. , , These are the weight matrices for the query vector, key vector, and value vector, respectively. The dimension of the key vector.

[0037] The dual-channel neural network model prediction method is as follows:

[0038] First, the time series characteristics Linear mapping to the query vector in the cross-modal attention mechanism Foreign trade characteristics Mapped to key vectors respectively Sum value vector Subsequently, the association score between the query vector and the key vector is calculated. The values ​​are then normalized using Softmax and used as weights for foreign trade features to determine the degree of participation of foreign trade features at the current time. Next, these foreign trade feature weights are compared with the value vector. Multiply by the product to calculate the foreign trade features of interest; finally, use the model parameter weights. and The original time series features and the key characteristics of foreign trade The final fusion feature is obtained by performing a weighted summation. The above-mentioned fusion features The input is fed into the decoder module, where the LSTM within the decoder performs temporal reconstruction and deep analysis on the complex information contained in the fused features, outputting a hidden state vector reflecting future load change trends. Finally, a fully connected layer performs a linear transformation on the hidden state vector, outputting the final predicted power load value for the future time. .

[0039] After adopting the technology proposed in this invention, the outward-oriented urban power load forecasting method according to the embodiments of this invention has the following advantages: it can adapt to changes in the international trade environment and accurately reflect its lagging effects, thereby improving the accuracy of power load forecasting results and helping to ensure the safe and stable operation of the power system and the reliable supply of electricity. Attached Figure Description

[0040] Figure 1 A flowchart of the outward-oriented urban power load forecasting method according to the present invention is shown. Detailed Implementation

[0041] Preferred embodiments of the invention will now be described with reference to the accompanying drawings. The following description, with reference to the accompanying drawings, is provided to aid in understanding exemplary embodiments of the invention as defined by the claims and their equivalents. It includes various specific details to aid understanding, but these should be considered exemplary only. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the invention. Furthermore, detailed descriptions of functions and structures well known in the art will be omitted to make the specification clearer and more concise.

[0042] like Figure 1 As shown, a method for forecasting electricity load in export-oriented cities that considers the lag in the transmission of foreign trade prosperity includes the following steps.

[0043] 1. Cleaning and alignment of multi-source heterogeneous data

[0044] The raw data falls into three categories, including: power load data. Let represent the maximum daily load of the region on day i, where i = 1, 2, ..., T (and so on). Meteorological data. These refer to the highest temperature, lowest temperature, average temperature, atmospheric pressure, relative humidity, average wind speed, rainfall, and snowfall in the region on day i, respectively. (Foreign trade data) These represent the Baltic Dry Index (daily), the RMB exchange rate of major exporting countries in the region (daily), the China Export Container Freight Index (weekly), the China Manufacturing New Export Orders Index (monthly), and the regional foreign trade cargo throughput (monthly), respectively.

[0045] (1) Outlier detection. For power load, box plots are used to remove outliers; for meteorological and foreign trade data, extreme fluctuations are retained to capture their impact on power load.

[0046] (2) Multi-scale alignment of foreign trade data. Considering that the short-term scheduling of power load is usually accurate to the day or even the hour, while the release frequency of foreign trade data is basically monthly, with a few being daily or weekly, we selected "daily frequency" as the basic frequency for modeling and adopted the strategy of "lowering the frequency to analyze the time delay and increasing the frequency to construct features".

[0047] Frequency reduction refers to directly using the release frequency of input indicators to perform statistical calculations on load. Specifically, for weekly release indicators (CCFI), the weekly maximum load is calculated during time lag analysis; for monthly release indicators (PMI, PT), the monthly maximum load is calculated by accumulating the daily maximum load. When performing time lag analysis, "load frequency reduction paired with foreign trade data" is used instead of "foreign trade data interpolation paired with load." Low-frequency data such as PMI and PT are interpolated into daily frequency data and load data for time lag analysis. This effectively avoids artificial smoothing noise and spurious trends introduced during linear or spline interpolation, ensuring that the calculated transmission lag reflects the true external shock transmission cycle, rather than the "pseudo-correlation" generated by interpolation, thereby improving the model's accuracy in identifying inflection points.

[0048] Upscaling refers to converting all indicators into daily frequency data. The specific operation method is as follows:

[0049] 1) Daily frequency data: Baltic Dry Index, RMB exchange rate, etc. remain unchanged and can be used directly.

[0050] 2) Weekly Data: Since the impact of freight rates on factories is a continuous process, this invention assumes that the market defaults to the current freight rate until the next price update. Therefore, the China Containerized Freight Index (CCFI) uses a forward-filling method. If the CCFI is released on Friday, the value for each day from Monday to Sunday of the previous week is equal to this value; if the CCFI has not yet been released, the value for each day from Monday to Sunday of the previous week and from Monday to Thursday of this week is equal to the CCFI released on Friday of the previous week.

[0051] 3) Monthly Frequency Data: The China Manufacturing New Export Order Index uses a forward-filling method. If the index is 50.1 released at the end of January, then the daily value from January 1st to January 31st is set to 50.1. If the index is not yet released for the current month, the index from the previous month is used. This "zero-order hold" method based on release time prevents the "future data leakage" problem caused by the traditional interpolation method using future data points to smooth the current point, ensuring that the temporal logic of the training model is consistent with the actual scheduling scenario, and significantly improving the practicality and stability of the model. For cumulative data such as regional foreign trade cargo throughput, this invention uses "daily average + forward filling," dividing the monthly foreign trade cargo throughput by the number of days in the month to obtain the daily foreign trade cargo throughput for the current month, eliminating the dimensional fluctuations caused by differences in the number of days in the month. When the monthly foreign trade cargo throughput is unknown, the data for the current month is not used; instead, the daily average foreign trade cargo throughput of the most recent known month is mapped. If a simple averaging method is used, the accuracy of subsequent load forecasting methods will not be high. Indicators such as the number of daily arriving ships that are proportional to the throughput of foreign trade goods can be introduced. By weighted decomposition, the monthly foreign trade goods throughput can be constructed into a daily frequency sequence to capture the instantaneous impact of short-term pulses in the international logistics market on the regional power load.

[0052]

[0053] in, It is the foreign trade cargo throughput on day i. The volume of foreign trade goods in this month, This represents the number of ships arriving on day i.

[0054] (3) Normalization. Due to the significant differences in the magnitudes of exchange rates, indices, and loads, data normalization is required. In this embodiment of the invention, the data reconstructed from features 2 (4) is normalized. For any original time series... The normalization process is as follows:

[0055]

[0056] in, This represents the value on day i after normalization of the original sequence. This represents the value of the original sequence on day i. Represents the minimum value in the sequence. This represents the maximum value in the sequence. The final output is a set of dimensionless, multidimensional time series vectors.

[0057] 2. Time Delay Analysis Based on the Maximum Information Coefficient Method

[0058] Considering that the traditional Pearson correlation coefficient can only capture linear correlations between variables, while the impact of foreign trade orders on electricity load often exhibits a "threshold effect" or "saturation effect," this invention employs the maximum information coefficient method to measure the transmission lag in order to identify this nonlinear relationship between external shocks and production electricity consumption. This solves the problem of linear models failing under scenarios of severe global economic fluctuations. The specific operation method is as follows:

[0059] (1) Set the maximum transmission lag period K. For foreign trade data, K can be set to 90 days = 3 months, because the cycle from signing the contract and preparing materials to electricity consumption for production in foreign trade orders is usually 1-3 months.

[0060] (2) Calculate the maximum information coefficient (MIC). For each input indicator, calculate the MIC at different transmission lag days. Below, its maximum information coefficient (MIC) with daily load. Taking BDI as an example, without lag, constructing data pair sequences. Calculate its maximum information coefficient When the propagation lag days k=1, construct a misaligned data pair sequence. Calculate its maximum information coefficient ,in This indicates that a 1-day lag in BDI affects the electricity load; and so on, until k=K, to calculate a set of MIC values. The specific calculation process for BDI and L's MIC is as follows:

[0061] 1) Assume two-dimensional variables BDI and L. Using the concept of entropy, we define mutual information between variables, which is the average reduction in uncertainty of L after being affected by BDI when BDI is known.

[0062] First, a resolution is given. The grid G ​​maps BDI and L to a two-dimensional scatter plot distributed within the grid. Then, the mutual information between BDI and L is calculated using a given partitioning scheme.

[0063]

[0064]

[0065] In the formula, Represents the mutual information value of two-dimensional variables. Let be the joint probability density function. and For edge density function, Indicates resolution as The maximum mutual information value under the grid G ​​partitioning.

[0066] 2) The maximum mutual information value under the single-resolution grid division method is normalized according to method 1(3) to obtain .

[0067] 3) Change the mesh resolution and repeat steps 1) and 2) to obtain the maximum mutual information value under different mesh sizes. .

[0068]

[0069] In the formula, Let BDI and L be the constraints for the mesh G during mesh generation, where B is a function of n, set according to general standards. .

[0070] (3) Determine the optimal conduction lag period The closer the MIC is to 1, the stronger the correlation between the two variables. The optimal transmission lag period of the input indicator is the number of days corresponding to the maximum MIC of the daily load under different transmission lag days, i.e. The k value corresponding to the maximum value in the middle.

[0071] (4) Feature Reconstruction. When inputting data into the prediction model, the foreign trade data is not the data for the current day, but rather... The data is used to incorporate the time difference between foreign trade factors and electricity demand into the construction of the prediction model. For weekly CCFI data and monthly PMI and PT data, the transmission lag weeks and months are determined directly using weekly and monthly data according to (1)-(3), and then converted into days, which are then input into the prediction model. or Data from 30 days is used to avoid errors caused by data alignment. The final output is a vector of each foreign trade indicator after time lag alignment. :

[0072]

[0073] In the formula, These are the Baltic Dry Index, the RMB exchange rate of major exporting countries in the region, the China Export Container Freight Index, the China Manufacturing New Export Orders Index, and the regional foreign trade cargo throughput transmission lag. The value corresponding to "day".

[0074] 3. Dual-channel neural network model prediction

[0075] This invention employs a "dual-channel feature fusion architecture," designing a time-series fluctuation channel and a foreign trade fluctuation channel, thus solving the problem that traditional single-channel models struggle to simultaneously consider both meteorological sensitivity and foreign trade changes. This architecture allows the model to primarily rely on the time-series fluctuation channel during sudden weather changes, while utilizing the foreign trade fluctuation channel to adjust the load baseline during severe external shocks. This ensures high prediction accuracy even when facing issues such as tariff adjustments and maritime crises. The specific design is as follows:

[0076] (1) Channel A: Temporal coding module based on CNN-BiLSTM. Constructing time series tensors. This includes historical electricity load data for T days, meteorological data (maximum temperature, minimum temperature, average temperature, atmospheric pressure, relative humidity, average wind speed, rainfall, snowfall), and date type features (weekday / holiday). First, a convolutional neural network (CNN) is used to process the input... Local feature extraction is performed to learn short-term local fluctuations in load and meteorological sequences, resulting in a convolutional output. Next, considering that CNNs cannot capture the periodicity, trends, and event-driven long-term dependencies of data, the convolution output is... The data is fed into a Bidirectional Long Short-Term Memory (BiLSTM) network, which utilizes forward and backward LSTM layers capable of acquiring sequence context associations to learn the daily, weekly, and yearly periodic dependencies between load and weather, ultimately outputting time-series features. .

[0077] (2) Channel B: Foreign trade coding module based on TCN. Construct a foreign trade indicator tensor aligned with time delay. This includes the Baltic Dry Index, RMB exchange rates of major exporting countries in the region, China's export container freight index, China's new export orders index for manufacturing, and the region's foreign trade cargo throughput. A Temporal Convolutional Network (TCN) is used to process the input. This study utilizes causal convolution and dilated convolution to extract the long-term evolution trend of foreign trade prosperity and outputs foreign trade characteristics. .

[0078] (3) Cross-modal feature fusion module based on attention mechanism. Cross-modal feature fusion integrates time series and foreign trade features. Although the two modules have different encoding methods, they both consider the features that affect the change of power load. By using foreign trade features to enhance the representation ability of time series features, the prediction of power load can be achieved.

[0079] Time-series characteristics and foreign trade characteristics are respectively and To match the tensor sizes of the two modal features, a query vector is obtained through linear mapping. Key vector Value vector The process of implementing the cross-modal attention mechanism is as follows:

[0080]

[0081] In the formula, The final feature vector after fusion. , These are the original feature preservation coefficients and cross-modal correction coefficients, respectively, which are automatically updated during model training. Softmax() is the probabilistic function. , , These are the weight matrices for the query vector, key vector, and value vector, respectively. The dimension of the key vector.

[0082] In the above process, firstly, the time series characteristics are... Linear mapping to the query vector in the cross-modal attention mechanism Foreign trade characteristics Mapped to key vectors respectively Sum value vector Subsequently, the association score between the query vector and the key vector (i.e., ...) is calculated. The weights are then normalized using Softmax and used as weights for foreign trade features to determine the degree of participation of foreign trade features at the current time. Next, these foreign trade feature weights are compared with the value vector. Multiply to calculate the foreign trade features of interest (i.e., the output of the attention mechanism); finally, use the model parameter weights. and The original time series features and the key characteristics of foreign trade The final fusion feature is obtained by performing a weighted summation. .

[0083] (4) Model prediction. The above-mentioned fused features... The input is fed into the decoder module, where the LSTM within the decoder performs temporal reconstruction and deep analysis on the complex information contained in the fused features, outputting a hidden state vector reflecting future load change trends. Finally, a fully connected layer performs a linear transformation on the hidden state vector, outputting the final predicted power load value for the future time. .

[0084] The present invention has been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

[0085] Through the above description of the embodiments, those skilled in the art can clearly understand that the present invention is feasible. Of course, the situations listed above are merely examples, and the present invention is not limited thereto. Those skilled in the art should understand that other variations or simplifications of the technical solutions according to the present invention can be appropriately applied to the present invention and should be included within the scope of the present invention.

Claims

1. A method for forecasting outward-oriented urban electricity load, characterized in that: First, the multi-source heterogeneous data is cleaned and aligned, including the following steps: (1) Collect raw data, including: power load data Meteorological data Foreign trade data ; (2) Outlier detection: For power load, box plots are used to remove outliers; for meteorological and foreign trade data, extreme fluctuations are retained to capture their impact on power load. (3) Multi-scale alignment of foreign trade data: Select daily frequency as the basic frequency for modeling, down-frequency analysis of time lag, and up-frequency construction of features; (4) Normalization process: For any original time series The data is normalized. Secondly, a time delay analysis based on the maximum information coefficient method is performed, including the following steps: (1) Set the maximum conduction lag time K; (2) Calculate the maximum information coefficient (MIC). For each input index, calculate the MIC under different transmission lag days. Below, its maximum information coefficient MIC with daily load; (3) Determine the optimal conduction lag period That is, the number of days corresponding to the maximum MIC of daily load under different transmission lag days, i.e. The k value corresponding to the maximum value in the middle; (4) Feature reconstruction: When inputting data into the prediction model, foreign trade data is not input for the current day, but rather... Data from the past day; Finally, the final future power load forecast is output through a dual-channel neural network model. The dual-channel neural network model includes a time-series fluctuation channel and a foreign trade fluctuation channel.

2. The outward-oriented urban power load forecasting method according to claim 1, characterized in that, The dual-channel neural network model includes: (1) Channel A: Temporal coding module based on CNN-BiLSTM; constructing time series tensors It includes historical electricity load data for T days, meteorological data, and date type features; the construction method involves, firstly, processing the input data through a convolutional neural network. Local feature extraction is performed to learn short-term local fluctuations in load and meteorological sequences, resulting in a convolutional output. Then, output the convolution. The data is fed into a bidirectional long short-term memory network, which utilizes its forward and backward LSTM layers to acquire sequence context associations. This network learns the daily, weekly, and yearly periodic dependencies between load and weather, ultimately outputting time-series features. ; (2) Channel B: Foreign trade coding module based on TCN; constructing a foreign trade indicator tensor aligned by time delay. It includes the Baltic Dry Index, the RMB exchange rate of major exporting countries in the region, the China Export Container Freight Index, the China Manufacturing New Export Orders Index, and the regional foreign trade cargo throughput; the construction method involves using a temporal convolutional network to process the input. This study utilizes causal convolution and dilated convolution to extract the long-term evolution trend of foreign trade prosperity and outputs foreign trade characteristics. ; (3) Cross-modal feature fusion module based on attention mechanism; cross-modal feature fusion integrates time series and foreign trade features, the time series features and foreign trade features are respectively and To match the tensor sizes of the two modal features, a query vector is obtained through linear mapping. Key vector Value vector The specific process of implementing the cross-modal attention mechanism is as follows: In the formula, The final feature vector after fusion. , These are the original feature preservation coefficients and cross-modal correction coefficients, respectively, which are automatically updated during model training. Softmax() is the probabilistic function. , , These are the weight matrices for the query vector, key vector, and value vector, respectively. The dimension of the key vector.

3. The outward-oriented urban power load forecasting method according to claim 2, characterized in that, The dual-channel neural network model prediction method is as follows: First, the time series characteristics Linear mapping to the query vector in the cross-modal attention mechanism Foreign trade characteristics Mapped to key vectors respectively Sum value vector Subsequently, the association score between the query vector and the key vector is calculated. The values ​​are then normalized using Softmax and used as weights for foreign trade features to determine the degree of participation of foreign trade features at the current time. Next, these foreign trade feature weights are compared with the value vector. Multiply by the product to calculate the foreign trade features of interest; finally, use the model parameter weights. and The original time series features and the key characteristics of foreign trade The final fusion feature is obtained by performing a weighted summation. The above-mentioned fusion features The input is fed into the decoder module, where the LSTM within the decoder performs temporal reconstruction and deep analysis on the complex information contained in the fused features, outputting a hidden state vector reflecting future load change trends. Finally, a fully connected layer performs a linear transformation on the hidden state vector, outputting the final predicted power load value for the future time. .

4. The outward-oriented urban power load forecasting method according to claim 1, characterized in that, The power load data , representing the daily maximum load of the region on day i, where i = 1, 2, ..., T; the meteorological data These refer to the highest temperature, lowest temperature, average temperature, atmospheric pressure, relative humidity, average wind speed, rainfall, and snowfall in the region on day i, respectively; the foreign trade data... These represent the Baltic Dry Index, the RMB exchange rate of major exporting countries in the region, the China Export Container Freight Index, the China Manufacturing New Export Orders Index, and the regional foreign trade cargo throughput, respectively.

5. The outward-oriented urban power load forecasting method according to claim 1, characterized in that, The frequency reduction analysis time lag is that the input indicators directly use the release frequency to perform statistical calculations on the load; for weekly release indicators, the weekly maximum load is calculated when analyzing the time lag; for monthly release indicators (PMI, PT), the monthly maximum load is calculated by accumulating the daily maximum load.

6. The outward-oriented urban power load forecasting method according to claim 1, characterized in that, The upsampling construction feature is that all indicators are converted into daily frequency data.

7. The outward-oriented urban power load forecasting method according to claim 6, characterized in that, To convert all indicators to daily frequency data, the procedure is as follows: Daily frequency data, including the Baltic Dry Index and the RMB exchange rate, remain unchanged and can be used directly. Weekly data, including the China Export Container Freight Index, uses a forward-filling method; Monthly frequency data, including the China Manufacturing New Export Orders Index, uses the forward-filling method; Cumulative data, including regional foreign trade cargo throughput, is averaged daily and filled forward. The monthly foreign trade cargo throughput is divided by the number of days in the month to obtain the daily foreign trade cargo throughput for that month, thus eliminating dimensional fluctuations caused by differences in the number of days in the month. When the monthly foreign trade cargo throughput is unknown, the data for that month is not used; instead, the average daily foreign trade cargo throughput of the most recent known month is mapped.

8. The outward-oriented urban power load forecasting method according to claim 7, characterized in that, The method for converting the regional foreign trade cargo throughput indicator into daily frequency data also includes: Introducing a daily vessel arrival count index that is directly proportional to the volume of foreign trade cargo, the monthly foreign trade cargo throughput is constructed into a daily frequency series through weighted decomposition. in, It is the foreign trade cargo throughput on day i. The volume of foreign trade goods in this month, This represents the number of ships arriving on day i.

9. The outward-oriented urban power load forecasting method according to claim 1, characterized in that, The normalization method is as follows: in, This represents the value on day i after normalization of the original sequence. This represents the value of the original sequence on day i. Represents the minimum value in the sequence. Represents the maximum value in the sequence; the final output is a set of dimensionless multidimensional time series vectors.

Images