As one of the main recharge sources in irrigation district, the change of precipitation have a significant impact on groundwater dynamics. From 1953 to 2010, the annual precipitation in Jinghuiqu irrigation district appeared a significantly falling trend, while evaporation is an unnotable increasing trend (Figure 1(a)). In the 1950s, the average annual precipitation and evaporation in the irrigation district were 551.8 mm and 1277.4 mm, however, after 1990s, the average annual precipitation has turned into 457.7 mm, which has been reduced by 17.1% compared with the 1950s, while the evaporation is 1297.3 mm with an increase of 1.6%. Taking R/S method to predict the future trend of the precipitation and evaporation [6], the result show that both of the Hurst coefficient H > 0.5, which indicates precipitation in Jinghuiqu irrigation district will continue to decrease and evaporation will continue to increase in a future period.

Mining or artificial groundwater recharge can change groundwater dynamics. With the increasing size and extent of human exploitation in groundwater, the impact of anthropogenic factors on the region's groundwater supply and drainage conditions becomes more and more prominent. As a typical large-scale irrigation district, farming activities in Jinghuiqu are very fierce, as well as the agricultural water demands is very large, which accounts for 90% of the total water consumption. Water supply source in Jinghuiqu irrigation district mainly

consist of the Jinghe river diversion and groundwater exploitation, Thus it can be seen that the proportion of the two parts water supply source plays an important role on the dynamic characteristics of groundwater in the irrigation area.

From 1977 to 2010, the irrigation intake water from canal head and the surface irrigation water use in the irrigation district both showed a falling trend (Figure 1(b) and Figure 1(c)). Before 1980s, the irrigation intake water from canal head can be more than 4 × 108 m³, but in the 1990s was only about 1.59 × 108 m³, which has been reduced by 62.5%; The second is surface water irrigation quantity, approximately from 1.82 × 108 m³ (1977) to 1.01 × 108 m³ (2010). As we know, infiltration of surface irrigation water is a major recharge source of the groundwater; therefore, we can infer that a significant reduction of the irrigation intake water, which has been leading to a substantial reduction in groundwater recharge, is a major cause of the severe decline in groundwater level in Jinghuiqu irrigation district in recent decades.

Groundwater pumping is the most direct way of groundwater discharge in irrigation district. At the same time, part of the extracted water in turn recharged the local groundwater by irrigation activities. With the increase of the irrigation water ratio of well and channel (Figure 1(d)), the way of surface water irrigation in the 1970s has gradually transformed into the current groundwater irrigation, which means that in the case of the extraction of groundwater has not changed dramatically, the groundwater discharge is relatively increased. Therefore, it can be inferred: It is the combined effects of the increasing groundwater pumping and decreasing irrigation recharge that destroyed the water balance and caused the groundwater level continuing to decline.

The above analysis shows that, in recent decades, the external environment of the groundwater system in Jinghuiqu irrigation district have undergone a significant change, comprehensive effects of climate change and human activity has resulted in the decline of groundwater level in the irrigation area. The decreased precipitation and increased evaporation, also the reduced amounts of groundwater recharge and the increased amounts of excretion all led to groundwater level dropping gradually, which dropped from 395.4 m to 383.6 m in nearly 34 years (Figure 1(e)). In Luqiao and Zhangbu district, two large dropdown hoppers have formed. Meanwhile, destroyed by hanging space, off pump and collapse, 90 percent of the wells built in the 1970s have been scrapped.

Table 1 lists the correlation coefficient between each factor and the groundwater depth, where S1-S7 is the groundwater depth, precipitation, evaporation, accumulative total groundwater exploitation, irrigation intake water from canal head, surface water irrigation quantity and irrigation water ratio of well and channel, respectively. As can be seen from the table, the accumulative total groundwater exploitation has a most closely related degree with the groundwater depth (Table 1). The exploitation of groundwater, as a dual identity of recharge and drainage of groundwater, plays the most prominent role on the groundwater level dynamics in Jinghuiqu irrigation district. Besides, the amount of surface water irrigation, as well as the proportion of groundwater and surface water supply, also closely linked with the groundwater water depth. Compared to human factors, the impact of precipitation and evaporation is relatively weak.

BP (Back Propagation) neural network is proposed by a team headed by Rumelhart and McCelland, which has the powerful ability to deal with non-linear interpolation to obtain the mathematical mapping reflecting the internal law of the experimental data [7] [8]. The prediction model contains three parts: input layer, hidden layer and output layer, a classical 3-layer BP net as Figure 2.

In the Figure 2, X_n is the input vector of n components, H_q is the hidden vector of q components, and Y_m is the output vector of m components (Figure 2). As the BP algorithm is very slow when calculating the derivatives of the weights and bias, the conjugate gradient back-propagation method was used to accelerate the computing efficiency. The main idea is to adjust the weights of the network to make total error minimum. In the forward process, the input signals are processed from the input layer to the output layer and the states of the layer's nodes can only influence the states of the next layer's nodes. At the output layer, the value of the output is compared with the anticipant value. If there is any error, the error will be returned along the quondam way, and the weight values of the nodes between layers are modified to reduce the error. So the error will be controlled in the range given in advance.

Based on the above analysis, groundwater level is affected by many factors. We select the annual precipitation, evaporation, groundwater exploitation, irrigation intake water from canal head, surface water irrigation quantity and irrigation water ratio of well and channel as the input vector, the groundwater level as the output vector. There were 43 samples data from 1977 to 2010. The samples data were divided into two parts: the first 33 samples data for training, the remaining 10 for predicting.

After data processing, a typical three-layered BP network is established. The final structure of BP network model is: 6-15-1, which means 6 input layer nodes, 15 hidden layer nodes and 1 output layer node. The non-li- near S function tansig is used in the hidden layer as the neuron transfer function and the purelin function as the output layer neuron prestige linear function. The specific training parameter: the training error is 0.001; the extensive training epoch is 8000; every 500 epochs demonstrates one time. In this model, when the number of hidden layer unit is 15, training results are very good.

Training and prediction results of model are shown in Table 2. The average absolute value of relative error of forecasting is 0.67%, which indicates that training and prediction values are good agreement with the actual value. Therefore, it can be concluded that the BP neural network model can reasonably predict the groundwater level in Jinghuiqu irrigation district.