基于BERT-BILSTM-CRF模型的电力行业事故文本智能分析*

为解决电力行业事故报告文本较长、语义复杂,难以进行有效文本识别问题,提出1种以BERT作为底层的预训练模型,并设计1种双重注意力机制编码器,结合BILSTM-CRF深度挖掘事故文本语义特征,从而实现文本智能分析。首先构建电力词典,通过对BERT预训练,进行BIO标注,然后引入BILSTM-CRF模型实现对文本标签智能分类,最后将该模型与现行其他4种深度学习模型进行对比。研究结果表明:该模型智能识别精确率、召回率及F1值(查准率)均达到约97%,较其他4种模型中效果最好的模型分别提高0.02,0.03,0.02。研究结果可为电力行业事故报告文本分析提供1种新思路。     

基于缓冲区重采样的LSTM林火预测模型*

为提高林火风险预测精度,挖掘地图上隐含的空间信息、时间序列上隐含的长期趋势和循环波动,提出1种基于缓冲区重采样的长短期记忆（LSTM）林火预测模型,选取15个与林火相关的影响因素,以方差膨胀因子为评价指标对其进行多重共线性检验,方差膨胀因子大于10的因素具有共线性,并采用信息增益率验证筛选结果的合理性。考虑到火灾的空间聚集特性,采用缓冲区分析与过采样相结合方法减少样本不均衡现象的影响,最终得到176 732条样本。对12个影响因素和研究时间段的火点建立LSTM预测模型,对森林火灾发生风险进行预测。研究结果表明:基于缓冲区重采样的LSTM林火预测模型有效考虑时空上隐含的信息,预测模型准确率为87.06%,特异性为97.99%,敏感度为76.12%,阳性预测率为97.43%,阴性预测率为80.41%,ROC曲线与AUC值均优于随机森林（RF）和支持向量机（SVM）这2种基准算法。维尔克松秩和检验发现,本文提出的模型与基准算法结果具有显著性差异。研究结果可为提高林火风险预测精度提供参考。     

2015年北京采暖季城市森林内外SO2浓度的时空变化特征

为探讨采暖季城市森林内外ρ(SO₂)动态变化及差异性,基于西山国家森林公园林内空气质量监测站数据,结合北京市环境保护监测中心植物园监测站的实时数据,分析2015年采暖季城市森林内外ρ(SO₂)变化和影响因素.结果表明:林内外ρ(SO₂)日变化基本呈双峰双谷型,在09:00-11:00和20:00-22:00左右达到高峰;采样期间ρ(SO₂)月变化呈不显著"V"型,最高值出现在1月,林内外分别为(25.8±9.2)和(31.7±23.4)μg/m3,最低值出现在11月,林内外分别为(19.0±5.2)和(13.0±11.2)μg/m3.林内ρ(SO₂)在1-3月低于林外,11-12月高于林外,林内ρ(SO₂)变化较林外平缓;气象条件对采暖季城市森林ρ(SO₂)变化有重要影响:降水对ρ(SO₂)消减效应明显,大风有驱散SO₂的作用,同时受风向影响;ρ(SO₂)和温度关系不显著(P=0.05,R < 0.40),但和空气相对湿度线性关系显著(α=0.05,Sig=0.00),林内受气象因素影响低于林外.研究显示,城市森林对气态污染物具有一定的缓冲、抵抗和吸收能力,因此应重视发展城市森林生态系统,充分发挥其生态效益,以提高城市大气环境质量.      

人为干扰对小兴安岭森林湿地土壤碳组分和酶活性的影响

以小兴安岭落叶林沼泽湿地为研究对象,对比研究皆伐地、火烧地和落叶松湿地土壤碳组分和酶活性.结果表明:(1)3种湿地类型土壤SOC含量表现为对照地显著高于皆伐地和火烧地(p0.05),在0~10 cm土层中皆伐地与对照地相比降低了58.38%,火烧地与对照地相比降低了61.96%.皆伐地与火烧地在不同土层的土壤中SOC含量差异不显著(p0.05);DOC含量对照地显著高于皆伐地和火烧地(p0.05).在0~10cm土层中火烧地DOC含量是皆伐地的1.99倍,两者差异不显著(p0.05);不同干扰方式对土壤MBC含量影响的差异主要表现在0~10 cm的土层土壤,对照地分别比皆伐地和火烧地高23.13%和95.79%(p0.05).在不同的土层中皆伐地MBC含量均大于火烧地,与SOC含量的变化趋势一致.(2)与对照地相比,皆伐地与火烧地0~10 cm土层的土壤蔗糖酶活性分别降低了45.59%和36.76%,脲酶活性显著降低了18.22%和55.69%(p0.05),蛋白酶活性降低了19.65%和17.34%,淀粉酶活性降低了6.29%和10.69%.土壤碳组分和酶活性在垂直方向上均表现为随着土层深度增加而减小.皆伐和火烧造成森林湿地土壤碳组分和酶活性降低.     

多层阻尼复合板常温贮存5年性能研究

目的分析多层阻尼复合板常温贮存5年的性能变化。方法对多层阻尼复合板进行常温下贮存5年的减振性能研究,通过贮存前后的正弦扫描试验和随机振动试验,获得贮存前后多层阻尼复合板的减振性能和模态变化,进行对比分析。结果阻尼复合板常温贮存5年后,径向模态有较大的下降,法向模态变化较少,组件上的加速度响应基本无变化。结论该多层阻尼复合板常温贮存5年后仍然满足要求。     

喀斯特常绿与落叶阔叶混交林过去50年来的碳循环模拟

通过对Forest-DNDC模型的植被和土壤参数本地化校准,以气象插值数据为输入,模拟了贵州省普定县高原型喀斯特次生常绿与落叶阔叶混交林1965~2014年的土壤、植被和生态系统碳循环特征。结果表明,与冷模拟和实测值相比,参数本地化校准后的模型能更准确地模拟春、秋、冬3个季节的土壤呼吸动态,而模拟的夏季土壤呼吸偏小;但统计检验指出,参数修订后的Forest-DNDC模型能够较好地模拟喀斯特森林土壤呼吸,降低了模拟误差,可用于喀斯特常绿与落叶阔叶混交林碳动态的模拟。进一步分析发现,1965~2014年喀斯特森林的碳通量除模拟早期的前3~4年急剧增加之外,随后总初级生产力(GPP)保持相对稳定,植物呼吸(Rplant)和生态系统呼吸(R_(ecosystem))随着森林发育而增加,土壤呼吸(R_(soil))减少,植被净初级生产力(NPP)呈迅速减小趋势;净生态系统碳交换量(NEE)亦较迅速下降,在2013年达到最低值-0.17 t C/ha,喀斯特森林由碳汇变为弱碳源。相关分析表明,年均温度和年降水对喀斯特常绿与落叶阔叶混交林的GPP和R_(soil)没有显著影响,但却显著影响NPP、R_(plant)、R_(ecosystem)和NEE。     

模拟氮沉降对重庆缙云山马尾松林土壤呼吸和酶活性的季节性影响

土壤酶参与土壤碳氮转化,同时土壤碳氮状况又是土壤酶活性的基础,而大气氮沉降通过影响土壤酶活性进而影响土壤CO_2释放.通过野外模拟试验,探讨不同氮沉降量对马尾松土壤呼吸和酶活性的影响,探索该区域马尾松土壤呼吸(Rs)与土壤温度(T)、土壤湿度(W)、Ure(脲酶)、Ive(转化酶)、CAT(过氧化氢酶)及ACP(酸性磷酸酶)的关系,为深入研究氮沉降对马尾松林森林生态系统的影响提供参考.2014年5月~2015年7月在缙云山马尾松林设置3个氮添加水平和一个无氮添加的对照处理:低氮[N_5,20 g·(m~2·a)~(-1)],中氮[N_(10),40 g·(m~2·a)~(-1)]、高氮[N_(15),60 g·(m~2·a)~(-1)]和对照[N0,0g·(m~2·a)~(-1)],每个处理量分4次,在每个季度开始各施1次,每个处理各9次重复,采用ACE(automated soil CO_2exchange station,UK)自动土壤呼吸监测系统分别对土壤呼吸、土壤温度和土壤湿度进行分析测定.结果表明:1土壤酶和土壤呼吸均具有明显的季节变化规律,各处理土壤呼吸均表现为夏季最高,其次是春季和秋季,最低为冬季,而各处理土壤酶活性则无一致的变化规律.2总体而言,氮沉降对土壤呼吸和酶活性均有抑制作用,且抑制程度随氮浓度增加而加强,但冬季氮沉降对马尾松林土壤呼吸有促进作用,春、夏、秋这3个季节氮沉降对Ure、Ive、CAT及ACP有抑制作用,而冬季氮沉降对4种土壤酶活性影响则存在差异.3逐步回归表明,无氮和低氮处理时,T、Ure和Ive对Rs的贡献较大,且随着T、Ure和Ive的增加,Rs也急剧增加;中氮处理时,T、Ure和CAT对Rs的贡献较大,Rs随着T、Ure和Ive的增加而增加;高氮处理时,Rs随着Ure的增加而降低,随着CAT和W的增加而增加.     

The multiple communities of low-carbon transition: an assessment of communities involved in forest carbon measurement

In recent writing on sociotechnical transitions theory communities are mostly conceived of as being local and place-based (“grassroots”). In this paper the implications for sociotechnical transitions theory of having multiple communities operating at different geographical scales, and with different objectives, are examined through a case study of low-carbon innovation in forests. The focus of analysis is the communities promoting sociotechnical innovations in the measurement of forest carbon. Innovation is being driven by the international United Nations climate policy initiative “Reducing Emissions from Deforestation and forest Degradation” (REDD+). The rise of REDD+ policy has prompted a flurry of activity in related scientific research, especially in the field of remote sensing. But other types of community are also actively positioning themselves as experts in forest carbon measurement: there are multiple communities at work, each with varying claims to innovation and expertise, from local (place-based) forest communities to international communities of foresters and forest ecologists. Recognition of the multiple communities operating within sociotechnical systems usefully draws our attention to the politics of innovation.     

Streamside Management Zones Compromised by Stream Crossings,Legacy Gullies,and Over‐Harvest in the Piedmont

Streamside management zone (SMZ) breakthroughs were identified and characterized to determine frequency and potential causes, in order to provide enhanced guidance for future water quality protection. Ten kilometers of SMZs were carefully examined for partial or complete breakthroughs. With partial breakthroughs the SMZ trapped sediment before it reached the stream, while complete breakthroughs appeared to have allowed sediment to have passed through with minimal restriction. A total of 41 breakthroughs occurred (33 complete, 8 partial) across 16 sites, averaging 1 complete breakthrough per 0.3 km of SMZ length. The most common complete breakthroughs were caused by stream crossings (42%), reactivation of legacy agricultural gullies (27%), and harvest related soil disturbances near/within SMZs (24%). Pearson correlations of site characteristics at breakthroughs indicated no strong relationships between breakthrough sites, representing the variable nature of these unique circumstances. Stream crossings are an intentional breakthrough for access purposes, but resulting environmental impacts can be reduced with best management practice implementation. Current recommendations for SMZs tend to work in most situations, yet further research is needed to identify causal factors and quantify breakthrough severity.     

Evaluating Alternative Temporal Survey Designs for Monitoring Wetland Area and Detecting Changes Over Time in California

Evaluation of wetland extent and changes in extent is a foundation of many wetland monitoring and assessment programs. Probabilistic sampling and mapping provides a cost‐effective alternative to comprehensive mapping for large geographic areas. One unresolved challenge for probabilistic or design‐based approaches is how best to monitor both status (e.g., extent at a single point in time) and trends (e.g., changes in extent over time) within a single monitoring program. Existing wetland status and trends (S&T) monitoring programs employ fixed sampling locations; however, theoretical evaluation and limited implementation in other landscape monitoring areas suggest that alternative designs could increase statistical efficiency and overall accuracy. In particular, designs that employ both fixed and nonfixed sampling locations (alternately termed permanent and temporary samples), termed sampling with partial replacement (SPR), are considered to efficiently and effectively balance monitoring current status with detection of trends. This study utilized simulated sampling to assess the performance of fixed sampling locations, SPR, and strictly nonfixed designs for monitoring wetland S&T over time. Modeled changes in wetland density over time were used as inputs for sampling simulations. In contrast to previous evaluations of SPR, the results of this study support the use of a fixed sampling design and show that SPR may underestimate both S&T.