Short-term load forecasting is the important foundation of formulating power generation
plans and adjusting operation modes for power dispatching departments, and its forecasting
accuracy marks the degree of power grid modernization. With the development of science and
technology in China, the power system is becoming more and more complex, the proportion
of new energy generation systems, electric vehicles, subways and other loads in the power
grid is gradually increasing, which makes the randomness and complexity of power load
gradually enhanced as well as increases the difficulty of short-term load forecasting.
Therefore, under the background of today's big data, how to use new methods to effectively
use data information to improve the accuracy of short-term load forecasting has important
research significance and value. Targeting improving the short-term load forecasting accuracy,
this paper improves the forecasting method from multiple perspectives of data processing,
forecasting model network structure and parameter optimization, establishing a high-precision
load forecasting model. The main research work is as follows:
First of all, a correlation analysis method of load influencing factors based on improved
fast correlation-based filter is proposed. The basic characteristics of power load and the
influencing factors such as temperature, humidity and electricity price are studied. Aiming at
the problems of low calculation efficiency and difficult determination of redundant
information in traditional feature selection algorithm, a method of analyzing load
auto-correlation and cross-correlation between variables and loads by using maximum
information coefficient is proposed, and the variables with strong correlation with load data
are screened. The redundant and invalid information in variables is eliminated by using
maximum information coefficient and fast correlation-based filter algorithm improved by
approximate Markov blanket, and finally the input variables of prediction model are determined.
Secondly, a GRU short-term load forecasting model based on Attention mechanism is
established. In order to solve the problems of information loss and low accuracy during GRU
processes long time series load data, an Attention layer is constructed between GRU network
hidden layer and fully connected layer, and thus the output proportion of GRU hidden layer is
redistributed by Attention mechanism, which enhances the sensitivity of GRU model to load
key information, effectively solves the forgetting defects when processing long time series
characteristics, and uses Adam algorithm to train the model. Finally, the open data set of an
Australian area and the load data of a China area are taken as practical examples to verify the
model. Compared with the prediction results of other methods, the experimental results show
that the prediction accuracy of GRU model improved by Attention mechanism is higher,
which proves the effectiveness of the improvement measures.
Finally, a short-term load forecasting method based on improved sparrow search
algorithm to optimize GRU-Attention is proposed. In view of the difficulties in selecting
parameters of GRU-Attention neural network and the low accuracy of the traditional
optimization algorithm, an improved sparrow optimization algorithm based on improved Sine
chaotic adaptive algorithm and Levy flight is proposed, and the parameters of the prediction
model are optimized by this algorithm, and the optimization results are assigned to
GRU-Attention model, which improves the prediction performance of GRU-Attention model.
The model is verified by the load data of different regions and compared with the prediction
results of the model before improvement. The results show that under different scale data sets
the proposed method has the lowest relative error and the best prediction stability, which
verifies the effectiveness and universality of the method.
谷歌
火狐
