随着互联网和电子商务平台的发展，目前我国已经成为最近几年全球最大的网络销售市场，网络购物这一消费模式也已经逐渐成为我国国民日常生活当中必不可少的一种生活方式。销售预测作为一种重要的应用性问题，能够为商家提供更加准确的以销定产，避免产品积压；合理管理产品库存，安排生产进度；及时的产品增补，避免供不应求等现象的产生。与传统的线下门店销售相比较，电子商务具有数据采集简易、数据处理迅速、数据量庞大等优势，可以利用这些优势更加便捷地进行商品销量预测，来为电商提供货备，营销策略等方面的调整策略。论文使用了某电商网店的历史销量数据，在Keras框架下搭建了三层LSTM神经网络模型，对比分析了CNN模型、传统的LSTM模型、ARMA-SVR组合模型，以及WaveNet-LSTM模型的预测的运算结果，得到新模型在预测准确度，训练效率等方面都有更大的优越性。 With the development of the Internet and e-commerce platforms, China has become the world’s largest online sales market in recent years. Online shopping, a consumption mode, has gradually become an indispensable way of life in the daily life of Chinese people. As an important application problem, sales forecast can provide more accurate sales target for merchants and avoid product overstocking. Manage product inventory reasonably and arrange production schedule; Timely product supplement, to avoid the phenomenon of short supply. Compared with the traditional offline store sales, e-commerce has the advantages of simple data collection, rapid data processing and huge data volume. We can make use of these advantages to forecast the sales volume of goods more quickly, so as to provide e-commerce with inventory, marketing strategy and other adjustment strategies. This paper uses the historical sales data of an e-commerce shop to build a three-layer LSTM network model under the framework of Keras. The calculation results of CNN model, traditional LSTM model, ARMA-SVR combined model and WaveNet-LSTM model are compared and analyzed, and the prediction accuracy, training efficiency and so on of the new model obtained have greater superiority.
销售预测，LSTM，电子商务, Sales Forecast LSTM Model E-Commerce摘要 - beplay安卓登录
With the development of the Internet and e-commerce platforms, China has become the world’s largest online sales market in recent years. Online shopping, a consumption mode, has gradually become an indispensable way of life in the daily life of Chinese people. As an important application problem, sales forecast can provide more accurate sales target for merchants and avoid product overstocking. Manage product inventory reasonably and arrange production schedule; Timely product supplement, to avoid the phenomenon of short supply. Compared with the traditional offline store sales, e-commerce has the advantages of simple data collection, rapid data processing and huge data volume. We can make use of these advantages to forecast the sales volume of goods more quickly, so as to provide e-commerce with inventory, marketing strategy and other adjustment strategies. This paper uses the historical sales data of an e-commerce shop to build a three-layer LSTM network model under the framework of Keras. The calculation results of CNN model, traditional LSTM model, ARMA-SVR combined model and WaveNet-LSTM model are compared and analyzed, and the prediction accuracy, training efficiency and so on of the new model obtained have greater superiority.
长短期神经网络是由循环神经网络发展而来，RNN的展开结构如图1所示。在时间序列预测中循环神经网络是使用频率相当高的一种网络，这一神经网络的具体结构示意图如下图2所示，这一图表当中对结构进行了全面的展示，在下图当中，这一神经网络所示有一个个紧密联系的神经元共同构成，形成了一个密闭的循环，而神经网络的右侧就是这一模型在计算过程当中的时间序列，其中 x t 时间输入， s t 代表输入时的这一时刻，而 o t 则代表输出的时刻，由此我们可以得出结论： o t 取决于 x t ，我们将数据按照时间顺序依次在神经网络的输入层植入，就可以获得与时间向关联的问题结果。但是最终决定实际结果的往往会与最终输入的数据直接相关，之前输入的数据会由于时间的逐渐流失而失去作用，所产生的效果也越来越低。针对时间长短的相关问题，提出了LSTM。
图1. RNN神经网络的展开结构
2.2. LSTM (长短期记忆神经网络)
相较于上述的RNN神经网络，长短期记忆神经网络在内部结构的设计方面更加精细，加入了更多的输入和输出单元，具体的单元结构图如下图2所示。其中 x t 输入时间， h t 表示状态值。图中三个大框分别表示不同时间点，神经细胞的实际状况，针对这一状况，本文将神经细胞的状态通过作图的模式展示出来，具体展示图如下图3所示 [11]。
图2. LSTM网络结构图
图3. 放大的神经元结构图
2.3. LSTM结构及功能实现
从图3可以看出，LSTM模型当 x t 和上一状态传递下来的 h t − 1 进行向量拼接乘以权重矩阵得到四个状态函数和输出分别如下：
i t = δ ( w i ∗ ( x t , h t − 1 + b i ) ) (1)
f t = δ ( w f ∗ ( x t , h t − 1 + b f ) ) (2)
o t = δ ( w i ∗ ( x t , h t − 1 + b o ) ) (3)
g t = tanh ( w g ∗ ( x t , h t − 1 + b g ) ) (4)
c t = c t − 1 ∗ f t + i t ∗ g t (5)
h t = o t ∗ tanh ( c t ) (6)
其中 i t , f t , o t 通过前面的sigmoid激活函数转换成0到1之间的数值，而构成门控单元，而 g t 则是将结果通过tanh激活函数转换成−1到1之间的数值作t时刻输入细胞的候选状态值。
想要对输入的数据 x t 进行筛选，并将其中对于后续预测有帮助作用的数据进行选择性的记忆时，这一模型的具体操作就是将需要记忆的数据进行输入，最终由上文当中的公式4进行计算，得出最终的记忆结果，对于数据的筛选功能则主要是通过上文当中的公式1实现的，两者的结果整合到一起传输到神经细胞的下一个状态当中 c t 。
2.3.3. 输出功能
LSTM输出当前的状态 h t 。主要是通过 o t 和tanh激活函数对 c t 进行缩放控制，使其具有长时间的记忆功能。
